Judicial Applications of Artificial Intelligence: Springer-Science+Business Media, B.V

JUDICIAL APPLICATIONS OF
ARTIFICIAL INTELLIGENCE
Edited by
GIOVANNI SARTOR
The Queen's University of Belfast
Northern Ireland, UK
and
KARL BRANTING
University of Wyoming
USA
Reprinted from Artificial Intelligence and Law 6(2-4) 1998
Springer-Science+Business Media, B.V.

A CLP. Catalogue record for this book is available from the Library of Congress.
ISBN 978-90-481-5136-3 ISBN 978-94-015-9010-5 (eBook)

DOI 10.1007/978-94-015-9010-5
Printed on acid-free paper
All Rights Reserved

©1998 Springer Science+Business Media Dordrecht
Originally published by Kluwer Academic Publishers in 1998
Softcover reprint of the hardcover 1st edition 1998
No part of the material protected by this copyright notice may be reproduced or

utilised in any form or by any means, electronic or mechanical,
including photocopying, recording or by any information storage and
retrieval system, without written permission from the copyright owner
CONTENTS
G. SARTOR and L. KARL BRANTING / Introduction: Judicial Appli-

cations of Artificial Intelligence 1
L. KARL BRANTING, JAMES C. LESTER and CHARLES B.
CALLAWAY / Automating Judicial Document Drafting: A Discourse-
Based Approach 7
URI J. SCHILD / Criminal Sentencing and Intelligent Decision Support 47
CYRUS TATA / The Application of Judicial Intelligence and 'Rules' to
Systems Supporting Discretionary Judicial Decision-Making 99
HENRY PRAKKEN and GIOVANNI SARTOR / Modelling Reasoning
with Precedents in a Formal Dialogue Game 127
PHILIP LEITH / The Judge and the Computer: How Best 'Decision
Support'? 185
MICHELE TARUFFO / Judicial Decisions and Artificial Intelligence 207
" ArtificialIntelligence and Law 6: 105-110, 1998. 105
Introduction: Judicial Applications of Artificial

Intelligence
G. SARTOR
Law Faculty, The Queen's University 0/ Belfast, Belfast B17 INN, UK
L. KARL BRANTING
Department a/Computer Science, University a/Wyoming, Laramie, WY 82071, USA
The history of artificial intelligence (AI) has been characterized by a steady expan-
sion of the aspects of human cognition and expertise amenable to computational
models. The earliest AI applications were in formal domains, such as theorem prov-
ing, that are relatively divorced from the complexity of ordinary human experience.
Progress in natural language processing, expert systems, planning, robotics, and
qualitative reasoning has extended the range of human experience and behavior
addressed by AI. This general trend also characterizes research on AI & law.
Early research addressed primarily routine legal domains, such as administrative
procedures, that involve the straightforward application of clear-cut rules to un-
controversial facts. Recent research, however, has focused increasingly on more
complex and challenging areas of the legal process typified by judicial decision-
making. No simple rule-chaining or pattern-matching algorithm can accurately
model judicial decision-making because the judiciary has the task of producing
reasonable and acceptable solutions in exactly those cases in which the facts, the
rules, or how they fit together are controversial.
Judicial decision-making is an area of daunting complexity, where highly so-
phisticated legal expertise merges with cognitive and emotional competence. Many
of the central concepts in the judicial application of the law - such as "justice",
"reasonable care", and "intent" - are deeply enmeshed in the fabric of human life.
Moreover, judicial reasoning combines diverse cognitive skills, such as assessing
facts, interpreting texts, making analogies, and engaging in dialectical interactions.
Besides its complexity, judicial decision-making is also characterized by its social
importance. It is, indeed, the most characteristic moment of legal experience. Indi-
vidual cases may involve important interests and deep feelings, and their solution
impacts upon the expectations of all legal actors and shapes their understanding of
the legal system.
These features of judicial activity justify a cautious approach. The hazards of
replacing judicial discretion with a rigid computer model can hardly be overesti-
[ 1]
106 G. SARTOR AND L. KARL BRANTING
mated (as stressed by Weizenbaum 1976, Gardner 1987, Berman & Hafner 1989,
among others). However, AI research projects in this field have consistently abjured
any attempt to usurp the discretionary reasoning of judges. Rather than aiming at
the impossible dream (or nightmare) of building an automatic judge, AI research
has aimed at developing practical tools to support judicial activities as well as new
analytical tools for understanding and modeling judicial decision-making.
1. Modeling judicial tasks
No form of legal reasoning seems to depend more heavily on uniquely human

abilities than the decision-making of a judge. Judicial decision-making requires
assessing the credibility of witnesses, evaluating the probative weight of evidence,
interpreting the meaning and intended effect of legal statutes and other normative
authorities and, especially in criminal cases, balancing mercy with justice. How
can AI contribute to a process that encompasses such a wide range of knowledge,
judgment, and experience?
The answer to this question, like the answer to the general AI problem of build-
ing intelligent artificial agents, is that one must begin by identifying the individual
tasks that collectively constitute the overall task of judicial problem solving. The
information-processing requirements of each of these individual tasks must then
be carefully analyzed, and computational methods identified that can satisfy these
information-processing requirements. When this analysis is complete, determining
appropriate algorithms and data structures to implement the required computational
methods is typically relatively straightforward. Finally, a conceptual presentation
of the knowledge involved in the tasks must be devised to mediate the input to, and
output from, the program and to facilitate knowledge acquisition. Often the most
valuable contribution of an AI project is the analysis of the underlying real-world
tasks and their information-processing requirements, since it is this analysis that
bridges the gap between expertise in the application domain and computer science
techniques.
Such an analysis is particularly important in developing automated systems for
as complex a process as judicial problem solving. A naive application of a given AI
algorithm, such as production rule chaining, neural nets, or decision-tree induction,
to a judicial task would be unlikely either to provide informative results or to be
accepted, were it divorced from the analysis, the nature of the task, its information-
processing requirements, or the necessary computational methods.
To face the forbidding objective of modeling judicial decision-making, AI &
law research must draw insights from many disciplines that have also studied judi-
cial decision-making, such as legal theory, legal procedure, psychology, sociology,
and organization theory. Legal theory, in particular, has traditionally dedicated the
largest part of its methodological analyses to judicial problem solving and offers
a rich palette of prescriptive and descriptive models. However, for several reasons
[2]
JUDICIAL APPLICATIONS OF ARTIFICIAL INTELLIGENCE 107
the contributions of these disciplines can only represent a starting point for an
AI-oriented analysis of the judicial process.
First, these disciplines do not provide the precision and the thoroughness that
is required of a computational model. In this first regard, AI & law research has
the potential to produce a new, more articulate, and rigorous representation of
the judicial practice. Such a representation, while supplying the background for
computing applications, can also provide a substantial contribution to the theory of
the judicial process and of judge-made law.
Second, the available models of judicial activities focus only on a few central
judicial activities, such as evaluating factual evidence and interpreting the law,
and disregard many ancillary judicial activities, such as producing documents and
accessing information. These ancillary tasks may in fact be the activities most
amenable to AI & law techniques.
Although the history of research in judicial applications of AI has been rela-
tively brief, AI & law has already obtained substantive results in both the directions
of research just mentioned. Contributions addressing the most central features of
judicial reasoning and judge-made law include the study of case-based reasoning.
In case-based reasoning, AI & law has provided new, powerful theoretical models
of aspects of judicial reasoning that integrate and develop the insights of legal
theory. Contributions to the analyses of ancillary activities include various projects
addressing preliminary or complementary judicial tasks (e.g., jurisdictional screen-
ing, drafting routine court documents, procedural tracking) or focusing on the
objective of helping pro se and other parties to successfully interact with the courts
(e.g., by assisting in the drafting of petitions and citations).
2. The papers in the special issue
Within this special issue we have tried to cover several different directions of cur-
rent research concerning AI and the judiciary, including descriptions of prototypes
and applications, formal analysis of the basic aspects of judicial reasoning, and
methodological analysis of judicial decision-making.
The opening paper by Karl Branting, James Lester, and Charles Callaway, en-
titled "Automating Judicial Document Drafting: A Discourse-Based Approach",
addresses a task that is ancillary to judicial decision-making but which has a sig-
nificant impact on the workload of judicial offices: the drafting of routine ancillary
judicial documents. Branting et al. propose a model of documents that makes ex-
plicit the goals that documents are intended to achieve and the stylistic conventions
to which they must conform. This model, termed the document grammar model,
can be used to automate the construction of new documents. Branting et al. describe
a representative class of judicial documents - appellate jurisdictional show-cause
orders - illustrate how show-cause orders can be represented in terms of a docu-
ment grammar, and describe an implementation of a prototype document planning
[3 ]
system that uses a document grammar to automate the drafting of new show-cause
orders.
The problem of the assistance to judges in the exercise of discretionary judg-
ment already addressed in contributions such as JEDA (Pethe et al. 1989) and
LawClerk (Branting 1993) is considered at length in Uri Schild's paper "Crimi-
nal Sentencing and Intelligent Decision Support". Schild's contribution focuses on
decision support in criminal cases, a problem which is highly controversial in many
countries. The paper first addresses the legal and political background of criminal
sentencing and the objective of sentencing policies. Schild argues that the primary
objective of sentencing systems should not be to impose some superficial measure
of uniformity, but rather to help the judge to balance uniformity (and fairness)
with the classical objectives of punishment (retribution, deterrence, prevention and
rehabilitation), in the framework of the policy of the legislator. Sentencing systems
should not substitute for the judge, but rather provide information or alternative
suggestions for human choices. Schild provides a critical review of various com-
puter systems for sentencing support, including algorithmic, rule-based, case-based
and model-based systems, and relates these systems to different sentencing poli-
cies. Finally, he describes a prototype case-based sentencing system and compares
case-based to statistical and rule-based approaches.
The problems of support to sentencing is also addressed by Cyrus Tata in the
paper entitled "The application of Judicial Intelligence and 'Rules' to Systems
Supporting Discretionary Judicial Decision-Making". Tata focuses primarily on
ways to assist the exercise of judicial discretion in penal cases, with special ref-
erence to a system intended to assist Scottish High Court Judges in the process
of criminal sentencing. After discussing various policies for sentencing, particu-
larly those reforms that are intended to produce uniformity and consistency by
restricting judicial discretion, he analyses various systems for sentencing support.
Tata focuses on systems that are intended to provide the user with information
about the range of penalties passed by the court for similar cases in the past.
For the selection of relevantly similar cases, he argues that classifications based
on doctrinal categories are inadequate. He advocates instead a "schematic holistic
representation", a broad classification reflecting the mental schemata used by the
judges for comparing criminal histories.
The article of Henry Prakken and Giovanni Sartor addresses the representation
of precedents and the application of case law. It builds upon previous research
(particularly the HYPO project by Kevin Ashley and Edwina Rissland) to provide
a general framework for legal precedent that integrates results from case based
reasoning and formal dialectics. Prakken and Sartor view judicial opinions as artic-
ulate (multi-argument, multi-step and multi-level) dialectical structures and strive
to preserve this articulation in their formal representation. They therefore represent
each precedent as a collection of arguments, any portion of which can be cited
in new disputes according to a dialectical protocol. The protocol provides heuris-
tics that replicate the basic non-deductive moves of case based reasoning (such as
[4]
JUDICIAL APPLICATIONS OF ARTIFICIAL INTELLIGENCE 109
analogizing and distinguishing a precedent). Finally, the paper presents a method

for assessing conflicts between precedents that generalizes HYPO's on-pointness
ordering and allows it to be combined with other choice criteria. The model de-
veloped by Prakken and Sartor is intended to provide both a formalized theory of
certain aspects of judicial law-making and a formal foundations for certain aspects
of computer programs for case-based reasoning.
The contribution by Philip Leith, entitled "The Judge and the Computer: How
Best 'Decision Support' ", both complements and contrasts with the views devel-
oped in the other papers of this special issue. While those papers adopt a cautious,
but positive and constructive, attitude towards AI applications in the judiciary do-
main, Leith stresses the limits of some AI technologies. He first discusses some
earlier expert systems (such as Prospector, Dendral and Mycin), arguing that they
were not accepted by the intended users because they were not based upon a
recognition of users' needs, function, and role. He further argues that AI & law
applications are bound to make the same mistake and therefore to being rejected
by their users. He stresses that discretion is valuable to the judges, so that every
attempt to reduce it through computers will be resisted by the judiciary. Moreover,
his view is that formalization (such as is required in the construction of a consis-
tent rule-based knowledge-base, sufficient to solve future cases) is fundamentally
incompatible with the judicial function in the most interesting areas of law. This
thesis is argued for from a rule-sceptical point of view, according to which legal
rules are only created by agents of the legal process (judges and barrister), sentenc-
ing is a social process characterized by the negotiation between those agents, in the
framework of social pressures and expectations, and under resource constraints.
The article by Michele Taruffo, which concludes this special issue, provides a
link between current jurisprudential research on judicial reasoning and of AI &
law. Taruffo views each case as a cluster of linked legal and factual issues, each
of which can be given different solutions. From this perspective the situation of
the court at the moment of its final decision is characterized by the existence of
several possible projects of decision and by the obligation of the court to choose
one of them as the best possible decision in that context. Taruffo observes that the
idea that the judge's reasoning could or should be completely reduced to simple
logical models is untenable. However, he observes that the failure of simple logical
models does not impair, but rather makes more valuable, AI & law research, which
has succeeded in going far beyond such simple models and which offers a wide and
growing inventory of analytical tools, some of which seem to fit judicial reasoning.
He focuses then on the automation of judicial discretion, distinguishing different
forms of discretion and discussing the contribution of AI to the treatment of each
of them. Finally, he addresses the problems of the logic of judicial justification,
where he points to the advantages that legal theory can derive from computational
dialectics.
[5]
3. Conclusion
The short history of AI for the judiciary, while emphasizing the difficult chal-
lenges that AI & law has to face, also displays the variety of solutions which
our discipline has so far provided. As a reply to those challenges, AI & law has
developed increasingly sophisticated models and techniques that address many of
the concerns of the critics of earlier AI models, including case-based reasoning,
formal dialectics, theory construction, neural networks, formal argumentation and
negotiation, intelligent document assembly, and tools for supporting discretional
decision-making. These achievements are useful and important for the judiciary,
because they provide a deeper and clearer understanding of some aspects of judicial
problem-solving, and an effective support to the judges and their collaborators.
Judges, squeezed between tightened budgets and increasing demands for jus-
tice, are desperately trying to maintain the quality of their decision-making process
while coping with time and resource limitations. Flexible AI tools for decision-
support may promote a sufficient degree of uniformity and efficiency in judicial
practice, while supporting a rational exercise of judicial discretion (and so possi-
bly help to prevent, for example, the draconian rigidity of compulsory sentencing
guidelines). In the same way, AI may help to reconcile flexibility, efficiency and ac-
curacy in complementary tasks, such as the drafting of various judicial documents.
In conclusion, we believe that the judiciary is in the early stages of a trans-
formation in which AI technology will make the judicial process faster, cheaper,
and more predictable without compromising the integrity of judges' discretionary
reasoning. We hope that the papers in this volume exemplify some of the directions
that this transformation will take.
References
Bennan, D. and Hafner, C (1989) 'The Potential of AI to Help Solve the Crisis in our Legal System',
CACM 32(8): 928-938.
Branting, K. (1993) 'An Issue-Oriented Approach to Judicial Document Assembly', In Proceedings
of the Fourth International Conference on Artificial Intelligence and Law, 228-235. New York:
ACM Press.
Gardner, A. (1987) An Artificial Intelligence Approach to Legal Reasoning. MIT Press: Cambridge
(MA).
Pethe, v.P., Rippey, CP., and Kale, L.v. (1989) 'A Specialized Expert System for Judicial Decision
Support', In Proceedings of the Second International Conference on Artificial Intelligence and
Law, 190-194. New York: ACM Press.
Weizenbaum, J. (1976) Computer Power and Human Reason. W.H. Freeman: San Francisco (CA).
[6]
ArtificialIntelligence and Law 6: 111-149, 1998.
111
© 1998 Kluwer Academic Publishers.
Automating Judicial Document Drafting: A

Discourse-Based Approach
L. KARL BRANTING
Department of Computer Science, University of ~oming, Box 3682, Laramie, WY 82071,
E-mail: [email protected]
JAMES C. LESTER and CHARLES B. CALLAWAY

Department of Computer Science, North Carolina State University, Box 8206, Raleigh, NC 27606,
E-mail: [email protected], [email protected]
Abstract. Document drafting is a central judicial problem-solving activity. Development of auto-

mated systems to assist judicial document drafting has been impeded by the absence of an explicit
model of (1) the connection between the document drafter's goals and the text intended to achieve
those goals, and (2) the rhetorical constraints expressing the stylistic and discourse conventions of
the document's genre. This paper proposes a model in which the drafter's goals and the stylistic and
discourse conventions are represented in a discourse structure consisting of a tree of illocutionary
and rhetorical operators with document text as leaves. A document grammar based on the discourse
structures of a representative set of documents can be used to synthesize a wide range of additional
documents from sets of case facts. The applicability of this model to a representative class of judi-
cial orders - jurisdictional show-cause orders - is demonstrated by illustrating (1) the analysis of
show-cause orders in terms of discourse structures, (2) the derivation of a document grammar from
discourse structures of two typical show-cause orders, and (3) the synthesis of a new show-cause
order from the document grammar.
1. Introduction
Legal problem solving subsumes a number of distinct tasks, including analyzing
the legal consequences of actual or hypothetical sequences of actions, argumenta-
tion, advising clients, planning transactions, and drafting legal documents. Legal
document drafting is an essential professional skill for attorneys and judges. In
the U.S., a significant portion of attorneys' workloads consists of drafting docu-
ments intended to precisely stipulate legal relationships such as wills, contracts,
and leases, and persuasive documents arising from litigation such as pleadings,
motions, and briefs.
Document drafting is a central activity of the judiciary. Judges' resolutions of
the disputes that come before them are generally embodied in written documents.
These documents can vary in complexity from brief memos to lengthy appellate
opinions. While judges have primary responsibility for judicial document drafting,
[7]
112 L. KARL BRANTING ET AL.
they are assisted by various judicial staff members including administrative and
secretarial staff and law clerks.
Two factors impose very high requirements for correctness and consistency on
judicial documents. First, the Anglo-American system embraces the doctrine of
stare decisis under which judicial decisions can be used as an authority to resolve
subsequent disputes. As a result, the impact of a document may extend far beyond
the parties whose dispute gave rise to the document. Second, all judicial decisions
and orders except those of the highest court in a given jurisdiction are subject to
review by higher courts. A party adversely affected by a judicial decision has a
strong incentive to discover any error or inconsistency in the document embodying
the decision, since such error or inconsistency could be used to attack the decision
in a higher court. Thus, high standards of correctness and consistency are essential
in judicial document drafting.
Document drafting can be viewed as a kind of configuration task in which
textual elements are selected and arranged to satisfy the goals of the drafter and
to conform to the stylistic conventions of the document genre. One source of com-
plexity in document drafting is the combinatorics of selection and configuration
decisions, which create large search spaces characteristic of most synthesis tasks.
However, a more fundamental reason for the difficulty of document drafting is that
the goals that documents are intended to achieve and the stylistic conventions to
which they must conform are seldom made explicit. An explicit representation of
these goals and conventions is essential to the development of automated tools to
assist in the document drafting process.
This paper proposes a model of documents that makes the underlying goals and
conventions explicit and uses this explicit theory to assist in the construction of new
documents. These goals and conventions are expressed as a dependency tree having
two types of discourse operators as interior nodes: illocutionary operators, which
express the goals that a document achieves; and rhetorical operators, which express
the stylistic conventions of the document's genre l We term this dependency tree
the discourse structure of the document. The illocutionary and rhetorical operators
capable of describing the discourse structures of a set of documents constitutes a
document grammar for the documents.
We have previously argued that representation of the discourse structure of
documents can facilitate (1) retrieval, interpretation, and adaptation of previous
documents, (2) maintenance of multi-generation documents, and (3) compari-
son of documents at a deeper level than mere surface text (Branting and Lester,
1996a; Branting and Lester, 1996b). In this paper we focus on the task of drafting
new documents using a document grammar derived from the discourse structures
of a set of documents representative of a given document genre.
1 An illocutionary operator is a speech act such as informing, requesting, warning, or promising.

A rhetorical operator is a discourse or coherence relation, such as exemplification, generalization,
sequence, or elaboration. See (Allen, 1987) for a more detailed discussion.
[8]
AUTOMATING JUDICIAL DOCUMENT DRAFTING: A DISCOURSE-BASED APPROACH 113
Section 2 describes a representative class of judicial documents - appellate

jurisdictional show-cause orders - and describes the potential benefits of au-
tomating their creation. Section 3 illustrates how the show-cause orders can be
represented in terms of discourse structures and describes informally how a docu-
ment grammar derived from these discourse structures can be used in the drafting
of new show-cause orders. Section 4 describes our implementation of a prototype
unification-based document planning system, the DOCU-PLANNER, and presents
details of the process whereby a document grammar can be used to draft a wide
range of related documents. Related work is discussed in Section 5, and Section 6
sets forth the scope of this approach and future research.
2. Judicial Document Drafting

Judicial decisions are expressed in judicial documents. In the Anglo-American
legal system, the most prominent judicial documents are appellate decisions. Ap-
pellate opinions typically contain a summary of the facts of the case, identification
of the issues of law raised in arguments by counsel for each of the parties, pro-
nouncement of the legal propositions supported by the controlling authorities, and
declaration of a decision that resolves the issues by applying the legal propositions
to the facts of the case (B ranting , 1993b). The complexity and individuality of
appellate opinions makes automated assistance for such documents far beyond the
scope of current technology. However, courts produce a number of other more rou-
tine documents having considerable stylistic and substantive consistency, including
various types of orders issued in response to motions or sua sponte. A single case
may give rise to numerous motions at both the trial and appellate levels relating
to, e.g., pleadings, discovery, time extensions, motions for dismissal or summary
judgment, or sanctions for violations of trial or appellate rules. The rulings of trial
and appellate courts on these motions typically take the form of orders of varying
degrees of length and complexity. Jurisdictional show-cause orders are typical of
such orders.
Jurisdictional show-cause orders are generally issued during jurisdictional
screening, a process of determining whether the requirements for an appeal have
been satisfied. Jurisdictional screening is typically performed at the earliest possi-
ble stage of an appeal to permit cases with jurisdictional defects to be recognized
as soon as possible. This minimizes unnecessary consumption of limited judicial
resources.
In this paper, discussion of appellate jurisdictional screening will focus on the
Colorado Court of Appeals, where one of the authors, Karl Branting, worked for
several years as a staff attorney. The Colorado Court of Appeals typically receives
over 100 new cases per month. Screening these appeals is too complex for clerical
personnel, but must instead be performed by a staff attorney. The staff attorney
examines the case file to determine whether the subject matter, finality, and time-
liness requirements for appellate jurisdiction have been met. If there appears to be
[9]
a jurisdictional defect, the staff attorney drafts a show-cause order that sets forth
the apparent defect and orders the appellant to rebut the defect within a fixed time
period or face dismissal of the appeal.
Figure 1 shows a typical show-cause order. This order identifies an apparent
defect - an untimely notice of appeal - and orders the appellant to show cause,
within 14 days, why the appeal should not therefore be dismissed. Show-cause
orders typify legal documents that are produced in relatively high volume (several
hundred per year), are complex enough to require drafting by an attorney, yet have
a high degree of stylistic and substantive consistency.
Currently, staff attorneys at the Colorado Court of Appeals draft show-cause
orders manually. One approach is for the staff attorney to reuse only those "boiler-
plate" text blocks that are common to all show-cause orders (e.g., "From the notice
of appeal filed by appellant ... "). Alternatively, a staff attorney may refer to a col-
lection of previous show-cause orders. The attorney can search this collection for
a previous order that involved jurisdictional defects similar to those in the current
case. The reusable language from the previous case may then be transcribed into
the current case, or the staff attorney may photocopy the previous order, cross out
the portions inapplicable to the new case, and write in portions specific to the new
case.
Manual drafting of show-cause orders has several clear disadvantages. First, the
process is very time-consuming and laborious. The first approach, which reuses
only the most general boilerplate language, entails repeated replication of drafting
effort and creates a high likelihood of inconsistent language. The second approach,
reuse of similar orders, depends on the staff attorney's ability to find and appro-
priately modify previous orders. This in tum depends on the attorney's ability to
understand the relevant similarities and differences between the goals that must be
achieved by the current show-cause order and the intentions underlying previous
orders.
The difficulty of drafting correct and consistent documents is exacerbated by
frequent personnel changes. For example, at the Colorado Court of Appeals, few
staff attorneys are willing to do jurisdictional screening for more than six months,
and many do screening for as little as three months. As a result, jurisdictional
screening is typically performed by attorneys with only limited experience drafting
show-cause orders.
Drafting even routine judicial documents, such as show-cause orders, is la-
borious and error-prone for an inexperienced drafter. A high volume of routine
orders may therefore constitute a significant drain on judicial resources even if
each individual order is relatively straightforward. It is widely recognized that
rising caseloads constitute one of the most pervasive problems confronting the
judicial system in the United States (Snellenburg, 1989). Technologies to enable
judges to use their time and expertise as efficiently as possible are therefore of
great potential importance to the judiciary. Automating the drafting of relatively
routine documents could make a significant contribution to judicial efficiency.
[10 ]
Colorado Court of Appeals Order

No. 87CA0514 Tr. Ct. No. 85CV269
STUART A. CANADA
Appellant
and
RODNEY T. WOOD, M.D., P.C., PENSION TRUST
Appellee
To: Stuart A. Canada and his attorneys, Mark J. Rubin
and Richard S. Strauss
From the notice of appeal filed by appellant and the
register of actions submitted by the clerk of the
district court, it appears that defendant is appealing
from both the trial court's order granting summary
judgment in favor of the plaintiff and the trial
court's subsequent order denying defendant's
C.R.C.P. 60 motion for relief from judgment. However,
it appears that the trial court's order granting
summary judgment in favor of the plaintiff was
entered February 9, 1987 and mailed to counsel of
record on February 10, 1987 and the notice of appeal
was filed on April 6, 1987. Furthermore, it appears
that the notice of appeal was due March 27, 1987.
Thus, it appears that the notice of appeal was not
timely as to the trial court's order granting summary
judgment in favor of the plaintiff. See C.A.R. 4(a).
IT IS THEREFORE ORDERED that the appellant shall show
cause, if any there is, in writing on or before
August 25, 1987 why this appeal should not be
partially dismissed with prejudice to the extent that
defendant seeks review of the trial court's order
granting summary judgment in favor of the plaintiff
for failure to file a timely notice of appeal.
BY THE COURT
Date: August 11, 1987
Copies to: Counsel of Record
Figure I. A typical show-cause order.
One approach to providing automated assistance in the drafting of show-cause

orders would be to design a set of templates for various show-cause orders, e.g.,
using WordPerfect macros. However, the wide variety of possible jurisdictional
defects and the even wider variety of factual situations that can give rise to juris-
dictional defects make devising an adequate collection of macros impracticable.
Moreover, even if a sufficient set of macros could be created, these macros would
[ 11 ]
present the user with an overwhelming number of choices, making selection of the
most appropriate macro unlikely.
We argue that the goals of accuracy, efficiency, and stylistic consistency are
best served by an approach to drafting routine legal documents that is based on
the discourse structure described in the next section. Our model of the document
drafting task is as follows:
Given:
• A set of relevant case facts.
• One or more illocutionary goals that the document is to achieve, e.g., estab-
lishing the prerequisites for dismissal.
• A document grammar which includes:
- illocutionary operators embodying the substantive legal rules governing
documents in the genre, and
- rhetorical operators embodying the stylistic and discourse conventions of
the genre.
Do:
• Planning. Find and instantiate a set of illocutionary operators that achieves
the document's illocutionary goals in terms of the given case facts and the
applicable set of rhetorical operators.
• Drafting. Synthesize a document that satisfies the illocutionary and rhetorical
operators.
There are several possible sources of the relevant case facts. One approach is
a conventional backward-chaining rule-based system, which would chain through
the applicable jurisdictional rules querying the user when necessary. Interactions
under this approach would be system-initiated. JEDA (Pethe et aI., 1989) and LAW
CLERK (B ranting , 1993a) illustrate system-initiated acquisition of case facts for
document drafting. An alternative, user-initiated, approach would use an electronic
form with entries corresponding to possible case values. A third alternative is a
mixed-initiative approach under which the user can directly provide case facts, ask
the system for examples, counter-examples and advice on answering questions, or
invoke an inference engine to help infer facts.3 The discussion below assumes only
that the relevant case facts have been obtained through one of these mechanisms,
but does not depend on the particular mechanism used.
The next section illustrates informally how the illocutionary and rhetorical goals
of show-cause orders can be represented by a discourse structure and how a docu-
ment grammar derived from this structure can be used to draft new documents. A
more detailed description of this process is set forth in Section 4, and a complete
document grammar for the examples in this paper appears in Appendix A.
3 This approach is used in the Jurisdictional Screening Assistant (JSA), a decision support system
for appellate jurisdiction screening under development at the University of Wyoming Department of
Computer Science.
[ 12]
3. Document Drafting Using Discourse Structures

3.1. ILLOCUTIONARY AND RHETORICAL STRUCTURE OF PERFORMATIVE
DOCUMENTS
Legal documents can serve a variety of illocutionary goals, including eliciting in-
formation, persuading, memorializing events such as reciprocal communications,
or accomplishing performative goals, such as creating or revoking legal relation-
ships. Judicial orders typically have a performative objective: they are intended to
define or alter legal relationships relevant to some controversy before the court.
There are generally three requirements that a performative judicial document,
such as an order or decision, must have to achieve the goal of defining or altering
a legal relationship. First, the document must find that some set of relevant facts
is present in the case. Second, the document must rule that one or more legal
propositions follow from applicable legal warrants under these facts. Finally, the
document must order some legal consequence justified by the legal propositions
under the given facts.4
For example, the show-cause order set forth in Figure 2 finds that the summary
judgment from which Appellant is appealing was granted on February 9, 1987
and mailed to Appellant on February 10, 1987, and that Appellant's notice of
appeal was filed on April 6, 1987. The show-cause order rules that the notice of
appeal was due on March 27, 1987 (45 days after notification of the judgment by
mailing). Finally, the Appellant is ordered to show why the findings or rulings are
not justified or suffer the sanction of dismissal with prejudice of the portion of the
appeal for which the notice of appeal was untimely.
The illocutionary goal of the show-cause order as a whole is to establish the
prerequisites for dismissal of the appeal. The findings and rulings required to es-
tablish the prerequisites for dismissal are determined by the legal rules governing
jurisdiction in the Court of Appeals. These are set forth in the Colorado Rules
of Civil Procedure (C.RC.P) and the Colorado Appellate Rules (C.A.R). For
example, C.A.R 4(a) provides that the notice of appeal "shall be filed with the
appellate court ... within forty-five days." The commencement of the 45 day period
is triggered by (1) "entry of the judgment or order appealed from" if the parties are
present at the time the judgment or order is announced, or (2) "the date of the
mailing of the notice" of judgment, if the notice is transmitted to the parties by
mail.
The illocutionary structure of the Canada v. Wood show-cause order is shown
on the left side of Figure 2. The top-level illocutionary goal is to establish
4 The findings of facts are analogous to the data in Toulmin's (Toulmin, 1958) model of argument.
The rulings of law are analogous to Toulmin's warrant and claim, since the rulings comprise both the
applicable legal authority and the conclusion that follows from applying the authority to the facts.
The additional element of performative judicial documents, the order, arises from courts' institu-
tional power to actually bring about changes in legal relations through documents of an appropriate
structure. The order is the final element necessary to bring about such a change.
[ 13 ]
No.I7CAOSI4 "fr. O. No. ISCV269
... ....,...
........
ROONEY T. WOOD, M.D., P.C., P£NS1OH TRUST
-)
Ur*('['hu5-fhrMI:;)
Unk (Show-a..e)
F~)_
8Y THE COUIIT
o.e.: A.... II. 1917
CopiaIlo; Quad or Reccrd
Figure 2. The illocutionary and rhetorical structure of Canada
the prerequlSltes for dismissal. This goal has two subgoals: to establish the
existence of a jurisdictional defect, and to order an appropriate sanction. The
relationship between an illocutionary goal, such as Establish(Untimely-notice-
ot-appeal), and its subgoals, Establish(Notice-ot-appeal-commencement-date),
Establish(Notice-ot -appeal-tiling-date), Establish(Notice-ot-appeal-due-date), and
Rule(Untimely-notice-ot-appeal), is expressed in an illocutionary operator.
The subtree underneath Establish(Jurisdictional-Detects) is similar to the goal
tree that would be generated by a rule-based system for determining the presence
of jurisdictional defects. In a conventional goal tree, legal rules would be used to
repeatedly decompose a goal establishing ajurisdictional defect into subgoals. Ulti-
mately, these subgoals would be grounded in the facts of the case. An illocutionary
structure differs from a conventional goal tree in that the leaf nodes are not limited
to case facts, but also include textual elements that satisfy illocutionary goals.
For example, under C.A.R. 4(a), determining that a notice of appeal is untimely
requires determining the date when the notice of appeal was filed. However, for
[14]
the document to achieve the illocutionary goal of establishing the date when the
notice of appeal was filed, the document must include text that makes a finding of
the filing date.
As shown in the left side of Figure 2, the illocutionary goal of establishing
a jurisdictional defect has as its sub goal establishing the orders being appealed
and establishing that the notice of appeal was untimely as to one of the orders.
Establishing untimeliness, in tum, has as subgoals: establishing the commencement
date of the time for filing a notice of appeal, establishing the due date of the notice
of appeal (45 days after commencement), establishing the actual filing date, and
ruling that the actual filing date was after the due date. The subgoals of these goals,
in tum, include: finding the judgment, mailing, and filing dates, ruling when the
date was due, and ruling that since the filing date was after the due date the notice
of appeal was untimely. The second subgoal for establishing the prerequisites for
dismissal is the show-cause order, which has as subgoals: ordering a time limit for
response, a sanction, and a rationale for the sanction.
In summary, the illocutionary goal structure expresses two kinds of information
essential to understanding the structure of a performative judicial document: the
goal dependencies among the applicable legal predicates (e.g., timeliness, method
of notification of jUdgment, and commencement of the time for filing a notice of ap-
peal); and the connection between performative text segments and the illocutionary
goals that they achieve.
Although the illocutionary goal structure represents information essential to
understanding the structure of performative judicial documents such as show-
cause orders, this structure is not per se sufficient to completely determine the
document's surface text. In general, the illocutionary goal structure does not
specify rhetorical features such as (1) the order of the textual elements that sat-
isfy various illocutionary goals, and (2) textual elements and stylistic constraints
imposed by the particular genre of the text, such as connective phrases and
other discourse cues. The right side of Figure 2 shows the rhetorical structure
of Canada v. Wood. The top-level goal is to Organize(Show-Cause-Order). The
subgoals are to provide frames for the caption (header), body, and footer of the
order. Within the body, the rhetorical structure includes discourse link features
characteristic of the show-cause order genre. Unlike the illocutionary structure,
the rhetorical structure is closely connected to the surface text of the document.
The relationship between rhetorical goals and their subgoals are expressed by
rhetorical operators. For example, the top-level rhetorical operator in Canada
permits the rhetorical goal Organize(show-cause-order) to be reduced to the goals
Frame(Header), Frame(body) and Frame(Footer). Together, the illocutionary and
rhetorical structures constitute the discourse structure of a document. 5
5 The discourse structure consists of a single dependency tree with both illocutionary and rhetor-
ical operators as interior nodes. For clarity, the illocutionary and rhetorical operators appearing in
the discourse structure are depicted separately in the left and right sides of Figure 2, respectively.
Discourse structure is discussed in more detail in Section 4.
[ 15]
Cokndo CGIIn CIt AA-II Order
--
No.9OC:Al1rT4 Tr. Q. No. 19DR221
........
TIX SUUnW.J(j~pro .. tppelIJn
It f'Idw ~ rhI: ttriI il not. ftna.I)JdpmeD:

~ it «- raend·..,.,.rtjcu_.:Don in wflicb
iI i ............... rdhi.fum-Ibrlheoourt
pronooncirW it to do in Older IO~J ddalnine the
RuII(AuthonI:y) . . orthe pM- irnoalved in the JIKlCOOdirw."

HrinrGl_Co. v.~640P.2d 1123(Coh 1912);
O.H. v. Pecple.192 a.,. 5042. 561 P.2d 331 (1965).
0..: Apl'iJ3.1990
Cap6a 110: Ccu-' 0{ RCIIXIId
Figure 3. The illocutionary and rhetorical structure of Kirkpatrick
In summary, the discourse structure of a document consists of illocutionary

and rhetorical operators that represent, respectively, (1) the connection between
the document drafter's goals and the text intended to achieve those goals, and (2)
the rhetorical constraints expressing the stylistic and discourse conventions of the
document's genre. The discourse structure grounds out in the text of the document.
3.2. USING DOCUMENT GRAMMARS FOR DRAFTING
The illocutionary and rhetorical operators necessary to construct the discourse

structures of a set of documents together constitute a document grammar for those
documents. To the extent that the document grammar is based on a representative
sample of the population of possible documents within the genre, the grammar will
be capable of generating a wide range of additional documents as well.
To illustrate this process informally, suppose that a document grammar has been
formalized to express the discourse structures of Canada and In re the Marriage
of Kirkpatrick, shown in Figure 3. The process of forming a document grammar
entails:
[ 16 ]
• Analyzing a representative set of documents to determine their illocutionary

and rhetorical structures,
• Extracting the illocutionary and rhetorical operators that appear in each
document's justification structure, and
• Generalizing the illocutionary and rhetorical operators.
Kirkpatrick's illocutionary structure differs from that of Canada in several

ways. First, the jurisdictional defect is a lack of finality rather than an untimely
notice of appeal. Second, a case that is not yet final may become final at some later
time, so the sanction for lack of finality is dismissal without prejudice, meaning that
the appeal can be filed again at some later date. This is in contrast to an untimely
notice of appeal, which can never become timely at a later date and for which the
appropriate sanction is therefore dismissal with prejudice. Finally, only one order
is being appealed in Kirkpatrick, so the sanction is not partial dismissal, as in
Canada, but complete dismissal.
Kirkpatrick also differs from Canada in its rhetorical structure. Kirkpatrick is a
domestic (i.e., divorce) case and therefore has a different caption than Canada, a
civil case. Moreover, Kirkpatrick's simpler illocutionary structure results in fewer
findings and rulings, so fewer link operators are required.
Suppose that a staff attorney is presented with a file for In re the Marriage of
Herbert W. Smythe and Catherine Smythe, a domestic case involving a summary
judgment entered on September 20, 1995 and mailed to the parties on September
22. Suppose that the appellant, Herbert Smythe, filed a notice of appeal on No-
vember 7, 1995, more than 45 days after commencement of the time period for
filing a notice of appeal. As discussed above, these facts might be gathered through
a decision-support system such as the Jurisdiction Screening Assistant or through
some alternative mechanism. Suppose that these facts, together with the docket
numbers on appeal and at trial, attorneys' names, etc., are provided to a document
drafting system. How could a document grammar for Kirkpatrick and Canada be
applied to these facts to draft an appropriate show-cause order?
The first step is to use the illocutionary operators to create a justification for
the goal Establish(Preqs-for-dismissal). This justification, shown on the left side of
Figure 4, is similar to the illocutionary structure in Canada in that for both cases
the jurisdictional defect is an untimely notice of appeal. The structure in Smythe
is simpler, however, because only a single order is being appealed. Moreover,
the existence of only a single appealed order means that sanction in Smythe is
complete dismissal, as in Kirkpatrick, rather than partial dismissal, as in Canada.
Accordingly, the illocutionary structure combines elements from both Kirkpatrick
and Canada.
The rhetorical structure of Smythe, shown on the right side of Figure 4, closely
resembles that of Kirkpatrick because both are domestic cases involving appeal of a
single order. Smythe's illocutionary and rhetorical structures are together sufficient
to determine the surface text of the order shown in Figure 4.
[17 ]
EabIiIlfl(Pleqt-for-4i._O No. 9SCA0431 Tc. Ct. No. 91CVOSI
BItab....Juri..icciorW-[)efedl)
L. . . Ob(Un. _ _ NOA'
Herbert: w. Smytbe
...
1b: HcdIert W. SmydIa. pro ...... _
--,
Rule(Unt_Iy-MOA) Sec C.A.R. "fa).
L
Link (SIlow-Ctue)
Ro""'..___ -
IT IS nmRI!RJRE ORDERED th.c: the appe... ..u
RUIe(A .. horiIy>/
-t""""'_-e.....,
Oeder(ll. . . . ...n.e..D.R'"
____~-l'-=~""
.r:::::::
............-
Oodw-o...+14o.y.) <---_-_ '0'-"-"--'_ _ _ _ _ _ _-1
BY THE COURT
OIderi'SIftCtioD.}
Dlte: [)reoembet4,l9IJS
ODderCUallmely-NOA-RMiDaae' Copiaco: c:o...lofRecurd
Figure 4. The illocutionary and rhetorical structure of Smythe
This example illustrates informally how a document grammar representing the

illocutionary and rhetorical operators underlying a set of representative documents
can be used to represent the illocutionary and rhetorical structures of new docu-
ments, which can in tum be used to generate the text of the document itself. A
formal model of a document grammar for Kirkpatrick and Canada and a unification
mechanism by which the text is realized from the resulting discourse structures is
described in the next section.
4. An Implemented Document Planner

To investigate the computational mechanisms required of automated document
planning for drafting judicial documents, we have designed and implemented a pro-
totype document planning system. Given the facts of a case, the Docu -PLANNER
[ 18]
AUTOMATING JUDICIAL DOCUMENT DRAFrING: A DISCOURSE-BASED APPROACH 123
automatically creates show-cause orders that identify apparent defects and issue
orders to appellants. The computational model on which the Docu - PLANNER is
based builds on a large body of work in computational linguistics on discourse
generation. Just as discourse generators produce multi-sentential texts by creating
hierarchical discourse structures, document planners can create documents in a
similar fashion. By planning a discourse structure and then drafting a document
that satisfies the illocutionary and rhetorical requirements dictated by this struc-
ture, the DOCU-PLANNER achieves the goals of accuracy, efficiency, and stylistic
consistency.
The DOCU-PLANNER is implemented in a unification-based formalism (El-
hadad, 1991). Its document grammar encodes the operators that are used to plan
documents. In contrast to syntactic grammars which specify the structure of well-
formed sentences, document grammars specify the structure of entire documents.
In particular, they specify how the facts of a case are used to create illocutionary
and rhetorical structures for a document to be generated for that case. Given a case,
the system creates a document in two phases:
• Document Planning: The system unifies the facts of the case with the document
grammar. This produces a discourse structure which is instantiated with the
specifics of the given case .
• Document Drafting: The system interprets the resulting illocutionary and
rhetorical specifications to create the final document in which the content,
rhetorical organization, and formatting are completely specified.
The DOCU-PLANNER can operate in one of two modes. In text mode, it creates
documents which are suitable for printing. In web mode, it produces documents
that have been formatted in hyper-text markup language and can be displayed
online with World Wide Web browsers.
This section is structured as follows. We first overview the theoretical foun-
dations of document planning by discussing computational models of discourse
planning. Next, we describe the DOCU-PLANNER'S implementation of docu-
ment grammars in the unification formalism. We then describe how the Docu-
PLANNER carries out document planning and document drafting. These are
illustrated with a sample document creation session in which the system creates
a show-cause order for a particular case.
4.1. FOUNDATIONS OF DISCOURSE GENERATION
The discourse-based approach to document planning builds on a strong founda-

tion laid by computational linguists in discourse generation. Although for many
years the primary focus of research in discourse structure was on accounting
for the coherence of expository or other communicative text for purposes of un-
derstanding, e.g., (Grosz and Sidner, 1986b; Hobbs, 1979), discourse generation
began to receive considerable attention beginning in the mid-1980s. Because doc-
ument construction is inherently a discourse generation task, the representations
[ 19]
that have been developed in the discourse generation community offer significant
insights for designers of document planning systems. Computational models of
discourse generation reason about the content and organization of knowledge to be
communicated in order to automatically construct multi-sentential text.
To produce discourse automatically, well-represented discourse knowledge is
crucial to the performance of discourse generators. Discourse knowledge is knowl-
edge about how to perform content determination (determining the content of dis-
course that is being constructed) and organization (determining the structure of the
discourse). The organizational aspect of discourse knowledge plays a particularly
important role in the construction of extended discourse. Discourse generation
is the process of applying discourse knowledge to produce multi-sentential or
multi-paragraph texts. We can distinguish three approaches to discourse genera-
tion: schema-based approaches, plan-based approaches, and hybrid approaches.
Beginning with work on schemata (McKeown, 1982; Paris, 1988), the field has
matured over the past decade and a half to produce top-down discourse planners
(Moore and Swartout, 1991; Suthers, 1991; Cawsey, 1992; Maybury, 1992; Hovy,
1993; Moore and Paris, 1993) and hybrid models (Suthers, 1991; Lester and Porter,
1997). We discuss each of these in tum.
The schema-based approach to discourse generation began with the pioneering
dissertation of McKeown (McKeown, 1982), in which she analyzed naturally oc-
curring texts to develop a set of schemata for describing concepts. Schemata, in this
context, are ATN-like structures that represent patterns of discourse. For example,
a schema for defining a concept includes instructions to identify its superclass, to
name its parts, and to list its attributes. Schemata contain rhetorical predicates,
e.g., "constituency," which names the parts of an object. Each rhetorical predicate
has an associated technique for extracting relevant propositions from a knowledge
base. Schemata order the rhetorical predicates, some of which are optional, some of
which can be repeated, and some of which can recursively invoke other schemata.
The schema-based approach has been very influential in discourse generation. Al-
though alternatives to schemata have since emerged, the schema-based approach
dominated the field for several years because schemata successfully capture many
aspects of discourse structure.
Schemata have been criticized because they lack flexibility. A top-down plan-
ning approach to discourse generation has been proposed to cope with this limited
flexibility. This approach, which has dominated the field for the past few years
(Moore and Swartout 1991; Suthers, 1991; Cawsey, 1992; Maybury, 1992; Hovy,
1993; Moore and Paris, 1993), can be traced to Appelt's work on planning referring
expressions (Appelt, L985), which in tum builds on earlier research on reasoning
about speech acts in a planning paradigm (Cohen and Perrault, 1979). Planners
offer a significant advantage over schema-based generators in that they can rea-
son about the structure, content, and goals of explanations, as opposed to merely
instantiating pre-existing plans embodied by schemata.
[20]
The operators of two seminal discourse planning systems are based on a theory
of discourse known as Rhetorical Structure Theory (RST) (Mann and Thompson,
1987). At the heart of RST is the assumption that multi-sentential texts have an
implicit structure that assists readers in assimilating the text's information. This
assumption suggests that writers - and discourse generators - should provide this
structure as they construct texts. RST was developed by analyzing a very large
corpus of texts. It consists of a small number of rhetorical relations, e.g., Back-
ground, Motivation, and Sequence. By recursively applying the relations, one can
parse a multi-sentential text into a tree where the leaves are clauses. Alternatively,
a discourse generator can employ a top-down planner with RST-like operators to
produce a multi-sentential text.
Perhaps the greatest problem faced by top-down planners is that their flexibility
comes at the cost of an enormous search space arising from the combinatorics of
mUltiple planning operators. The relative merits of schemata and top-down plan-
ners have been heavily debated. One conclusion is that schemata are best viewed as
"fossilized discourse structures" that represented previously compiled plans (Hovy,
1993). To respond to a number of problems in discourse generation, Suthers devel-
oped a sophisticated hybrid approach that includes planning techniques as well
as plan critics, simulation models, reorganization methods, and graph traversal
(Suthers, 1991). By assembling these diverse mechanisms into a single architec-
ture, he demonstrates how the complexities of discourse planning can be dealt
with in a coherent framework. Lester and Porter developed the hybrid approach
of explanation design packages (EDPs) for KNIGHT, a robust discourse generator
for large-scale knowledge bases (Lester and Porter, 1996; Lester and Porter, 1997).
KNIGHT'S EDPs, which constitute a schema-like programming language for "dis-
course knowledge engineers," combine a hierarchical frame-based representation
with embedded procedural constructs for knowledge-base access.
4.2. REPRESENTING DOCUMENT PLANNING KNOWLEDGE

Document planners can build on the large body of work in discourse generation.
Because content determination problems and organization problems are analogous,
many of the solutions that have been developed in discourse generation can be
adapted for document planning. Discourse generators have been studied in genres
that include expository texts, editorial texts, and advisory texts, inappropriate for
document planners. Unlike expository texts, few documents have an exclusively
communicative purpose. Rather, like editorial and advisory texts, documents are
frequently concerned with peifonnative utterances, a designation proposed by J.L.
Austin (Austin, 1962) and later elaborated by others in speech act theory (Grice,
1975; Searle, 1969). Because individuals and institutions frequently draft docu-
ments to accomplish performative goals, such as creating or revoking legal, social,
or institutional relationships, or eliciting information, representing the illocutionary
structure of documents is critical.
[21 ]
Schema-based approaches and their derivatives, such as explanation design

packages, have been most successful in producing expository texts. However,
because schemata do not record intentional information and because document
planners must be able to reason about the content of documents, a plan-based
approach to representing the illocutionary structure underlying documents is more
promising. Moore has observed that, "Any approach to discourse structure that
relies solely on rhetorical relations or predicates and does not explicitly encode
information about intentions is inadequate for handling dialogues" (Moore, 1995).
For precisely the same reason, a pure schema-based approach that omits intentional
knowledge would fare poorly in a document planning system.
Despite the great importance of illocutionary knowledge, the rhetorical structure
of documents is equally critical. For example, conforming to the precise structure
and formatting conventions of court documents is indispensable to the success of
document planners for judicial applications. Hence, we must encode rhetorical
knowledge as well, perhaps in a schema-based formalism. However, adopting a
hybrid model that employs two formalisms - a plan-based approach for illocu-
tionary knowledge and a schema-based approach for rhetorical knowledge - is
cumbersome. We therefore opt for a uniform approach and are presented with two
alternatives: we can either embed illocutionary and rhetorical knowledge in the
same operators - to some extent, this is the RST approach - or we can somehow
decouple them. Decoupling offers two important advantages:
• Increased fidelity of representation: Decoupling enables the illocutionary and
rhetorical theories to reflect the reality that, in many document planning tasks,
the illocutionary structure and rhetorical structure do not stand in a subsump-
tive relationship. This phenomenon seems to have arisen because the rhetorical
structure of documents has taken on a life of its own, spawning discourse
organization conventions that are sometimes orthogonal to the illocutionary
structure .
• Increased ease of inspection: Decoupling the structures permits users to view
either the full illocutionary structure or the rhetorical structure in isolation from
the other. If the two sets of operators co-exist separately, users can request an
illocutionary view of the document separate from the rhetorical structure, and
vice versa.
The discourse-bas.ed approach encodes knowledge about illocutionary and
rhetorical structures in a uniform, plan-based representation. A document gram-
mar, which encodes illocutionary and rhetorical operators, can be represented in
a functional unification grammar formalism (Kay, 1979). Given the facts of a spe-
cific case, the document planner sub-goals on the illocutionary operators to create
an illocutionary structure for the document while it sub-goals on the rhetorical
operators to create a rhetorical structure for the document.
Illocutionary operators represent the top-down decomposition of intentional
goals into sub-goals. For example, the goal of the operator
[22 ]
Establish(Jurisdictional-Defect) ===*
Establish(Appeals)
At-Least-One [ Establish(Untimely-Notice-Of-Appeal),
Establish(Subject-Defect),
Establish(Finality-Defect) ]
is to establish a jurisdictional defect. It accomplishes this by attempting to establish
what orders are being appealed from (i.e., Establish(Appeals)) and then, for each
such order, attempting to establish as many of the following as possible: untimely
notice of appeal, a subject defect, and/or a finality defect. If it is able to achieve the
goal of establishing appeals as well as at least one of the other three sub-goals, it
succeeds; otherwise it fails.
To further illustrate, consider the operator that establishes an untimely notice of
appeal:
Establish(Untimely-Notice-Ot-Appeal) ===*
IF <judgment-date>
{ Establish(NOA-Commencement)
Establish( <filing-date>, Notice-Ot-Appeal)
Establish( <due-date>, Notice-Of-Appeal)
Rule(Untimely-Notice-Of-Appeal) }
Five aspects of this operator are noteworthy. First, the goal of the operator appears
as a sub-goal in the preceding operator. As a result, when the system attempts
to satisfy the first operator which establishes a jurisdictional defect, the second
operator is invoked to establish an untimely notice of appeal. Second, the system
inspects its knowledge of the case to determine if there is a judgment date. If there
is one, it proceeds to the sub-goals; if there is not, it fails. Third, the sub-goals are
conjunctive: all of them must be achieved for the operator to succeed. The sub-
goal Rule(Untimely-Notice-Ot-Appeal) succeeds only if the <filing-date> is later
than the <due-date> Fourth, just as in logic programming, variable bindings must
be made consistently for the goal to be achieved. Finally, text-emission goals are
present in the Rule goal. If the illocutionary plan containing an instantiated version
of this operator ultimately succeeds, then text emission will occur.
Rhetorical operators represent the top-down decomposition of textual organiza-
tion goals into sub-goals. For example, the operator
Organize(Show-Cause-Order) ===* Frame(Header)
Frame(8ody)
Frame(Footer)
organizes the global structure of the major sections of documents. Some rhetorical
operators impose an organization on a particular section, e.g.,
Frame(80dy) ===* Link(Preamble)
Link(However-Phrase)
Link(Furthermore-Phrase)
Link(Thus-Phrase)
Link(Show-Cause)
[23 ]
In addition to the illocutionary and rhetorical operators, the document grammar

also includes organizational constraints that specify the interleaving of operators
that is required to produce the final instantiated discourse structure. For example,
the constraint
Sequence(Preqs-For-Dismissal) = Left-To-Right [ Frame(Header),
Link(Preamble ),
Establish(Jurisdictional-Defects),
Order( Show-Cause),
Frame(Footer) ]
specifies how achieving the illocutionary goals of establishing jurisdictional de-

fects and ordering the appellant to show cause should be interleaved with the
rhetorical subgoals of constructing a a header, preamble, and footer.
Our document grammar is organized into operator modules, each of which
contains a collection of operators that perform a similar function. The Docu-
PLANNER currently contains 7 modules with 85 operators (Figure 5). The Es-
tablishment, Ordering, Findings, and Rulings modules contain the illocutionary
operators. The Organizational and Constraint and Framing and Linking modules
contain the rhetorical operators. The Informing module contains operators that emit
text; these are invoked as sub-goals of both illocutionary and rhetorical operators.
Modularizing the document grammar in this fashion facilitates construction of new
operators and maintenance of existing operators.
Our operators are represented in a functional unification formalism. For ex-
ample, the four operators discussed above are represented as shown in Figure 6.
The syntax of the fOlmalism is that of functional descriptions (Elhadad, 1991),
which is defined recursively: a functional description consists of a list of pairs,
where the second item in each pair is either atomic or is itself a functional descrip-
tion. Collectively, the functional descriptions for the illocutionary and rhetorical
operators define the document grammar. Employing the unification formalism sig-
nificantly increases the planner's flexibility by enabling it to create documents for
an enormous variety of case facts.
4.3. DOCUMENT CREATION

The DOCU-PLANNER (Figure 7) is a unification-based implementation of a plan-
ning approach to document construction. Given the facts of an appellate case,
the DOCU-PLANNER constructs show-cause orders with the appropriate content,
organization, and stylization. It consists of three components:
• Document Grammar: Encodes the illocutionary and rhetorical operators.
• Document Planner: Constructs discourse structures (linked and instantiated
illocutionary and rhetorical operators).
• Document Drafter: Creates completed documents by traversing discourse
structures and emitting formatted text.
[24 ]
AUTOMATING JUDICIAL DOCUMENT DRAFfING: A DISCOURSE-BASED APPROACH 129
(1) The Establishment Module

Function: Creates backbone of illocutionary structure
Type: Illocutionary
Example: Establish-Untimely-Notice-Of-Appeal
Number of operators: 12
(2) The Ordering Module
Function: Enunciates performative text
Type: Illocutionary
Example: Sanction
Number of operators: lO
(3) The Findings Module
Function: Makes factual findings
Type: Illocutionary
Example: Judgment
(4) The Rulings Module
Function: Makes legal ruling
Type: Illocutionary
Example: Non-Final-Order
(5) The Organizational and Constraint Module
Function: Imposes rhetorical organization
Type: Rhetorical
Example: Sequence (Show-Cause)
Number of operators (and constraints): 8
(6) The Framing and Linking Module
Functions: Creates boiler plate text including header and footer
Creates linking phrases
Creates specialized formatting directives
Type: Rhetorical
Example: Frame (Header)
(7) The Informing Module
Functions: Emits and conjoins pre-computed text segments
Creates inter-paragraph formatting directives
Type: Text Emission
Example: Inform-Appeal-Filed
Figure 5. Operator modules of the DOCU-PLANNER's document grammar
[25 ]
((type jurisdictional-defect)
(untimely-defect ((alt untimely-defect
(((cat establish)
(type untimely-noa))
((cat stop)
(untimely-noa? no) )))))
(subject-defect ((alt subject-defect
(((cat establish)
(type subject-matter-defect))
((cat stop)
(subject-matter-defect? no))))))
(finality-defect ((alt finality-defect
(((cat establish)
(type finality-defect))
((cat stop)
(finality-defect? no))) ))))
((type untimely-noa)
(date-later-than #(external due-date))
(date-later-than yes)
(e-due-date ((cat establish)
(type noa-due-date)))
(e-filing-date ((cat establish)
(type noa-filing-date)))
(and ((cat text) (lex "and")))
(f-filing-date ((cat find)
(type noa-filing-date)))
(furthermore ((cat text) (lex "Furthermore, it appears that")))
(f-due-date ((cat find)
(type noa-due-date)))
(thus ((cat text) (lex "Thus, it appears that")))
(rule-untimely ((cat rule)
(type untimely-noa)))
(authority iauthority}))
(pattern (untimely-defect subject-defect finality-defect))
(pattern (appeal-filed filing-date end-sent)
Figure 6. Unification formalism of illocutionary and rhetorical operators
[26 ]
Case - - -___r.f Document Planner t-------....,

Facts
lllocutionary Rhetorical
Structure Structure
I I
IFinal
Document
Document Drafter
--
Figure 7. The DOCU-PLANNER'S architecture
Given the specifics of a particular case, the document planner backchains on the
illocutionary operators in a problem-decomposition fashion to construct the evolv-
ing document's illocutionary structure. Similarly, it backchains on the rhetorical
operators to construct the document's rhetorical structure. Both of these tasks are
accomplished simultaneously through unification of the representation of the case
facts with the document grammar. The net result of this computation is a discourse
structure in which the operators defining the illocutionary and rhetorical structures
are fully instantiated and linked together through variable bindings. Many nodes in
these structures specify the production of text segments and formatting directives.
Next, the document drafter conducts a pre-order traversal of the discourse structure
produced by the document planner. It then concatenates the text obtained from this
traversal and embeds formatting directives (which were also specified in the dis-
course structure) in the concatenated text. Finally, it interprets the resulting linear
structure, thereby creating the completed document.
The DOCU-PLANNER is implemented with FUF (Functional Unification For-
malism), a robust unification environment developed at Columbia University
(Elhadad, 1991; Elhadad, 1992). FUF is itself implemented in Lisp, as is the Docu-
PLANNER'S drafting system. The entire system runs in Harlequin Lisp on a DEC
Alpha. Document creation is accomplished very quickly. Given the case facts, the
typical amount of time to create an appellate jurisdictional show cause order is less
than one second.
To illustrate the DOCU-PLANNER's behavior, consider Smythe vs. Smythe. Re-
call that in Smythe Herbert Smythe filed a notice of appeal on November 7, 1995 in
[27 ]
Case: Smythe
Notification-Method: Mailing
Appellant: Herbert W. Smythe
Appellee: Catherine Smythe
Substantive-Legal-i'l.rea: Domestic
Judgmen t - Da t\~ :
Mont.h: 9
Day: 20
Year 1995
Filing-Date:
Month: 11
Day: 7
Year 1995
Mailing-Date:
Month: 9
Day: 22
Year 1995
Order-Date:
Month: 12
Day: 4
Year 1995
Case-Number: No. 95CA0437
Order-Number: Tr. Ct. No. 91CV051
Aut.hority: See C.A.R. 4(a).
Figure 8. Representation of case facts of Smythe vs. Smythe
response to a judgment that was entered on September 20, 1995. The summary
judgment for this domestic case was mailed on September 22. The DOeD-
PLANNER's representation of the facts of Smythe, including the cited authority,
the appellant, and the appellee, are represented in Figure 8.
Given these facts, the document planner first creates a discourse structure for
a show-cause order by unifying the input representation with the document gram-
mar. By unifying the top-level illocutionary operators with the input data, then
backchaining on these instantiated operators and repeating this process recursively,
the document planner constructs a fully instantiated illocutionary structure. Begin-
ning with the goal of establishing the prerequisites for dismissal, which is included
in the findings of the case, the document planner posts two sub-goals, establishing
jurisdictional defects and ordering the appellant to show cause. It then attempts to
achieve each of these in tum. To establish jurisdictional defects, the planner first
attempts to establish an untimely notice of appeal. To accomplish this sub-goal, it
attempts to establish that the due date for the notice of appeal preceded the filing
[28 ]
Establish~s-for-dismissaJ)
Establish(Jurisdictional-Defects)
L E6tablish(Untimely-NOA)
Establish(NOA-Commencement)
F"md(NOA-Commencement.
Judgement-Date)
F"md(F"!led-by-MaiI,
Mailing-Date)
E6tablish(NOA-F"!ling-Date)
L___ v...
F"!ling-Date)
E6tablish(NOA-Due-Date)
l Vmd(_-"'"
Due-Date = Mailing-Date + 45 Days)
Rule(Untimely-NOA)
1-
L
Rule(Untimely-Appea1)
Rule(Authority)
OrdeI(Show-Cause)
Order(Response-Due-Date =
Order-Date + 14 Days)
- OrdeI(Sandion)
Order(Untimely-NOA-Rationa1e)
Figure 9. The illocutionary structure constructed for Smythe
[29]
Organize(Show-Cause-Order)
Frame (Header)
Frame(Body)
Link(Preamble)
Link (Furthennore-Phrase)
Link (Show-Cause)
Frame(Footer)
Figure 10. The rhewrical structure of Smythe.
date of the appeal. By analyzing the date information contained in the facts of the
case, it determines that the goal is satisfied.
Next, it posts four sub-goals which, if conjunctively satisfied, will achieve
the goal of demonstrating that the appeal was untimely: (1) establish the date of
commencement for the notice of appeal (NOA); (2) establish the filing date of
the NOA; (3) establish the due date for the NOA; and (4) make a ruling that the
NOA was untimely. To achieve subgoal (1), it first makes a finding of the date of
commencement based on the information in the case data. It then attempts to make
a finding that the commencement date is the same as the judgment date, which is
stated in the case data. However, in order for the judgment date to be the same
as the commencement date, the appellant must have been present at judgment.
The case data contradicts this proposition, thereby causing this sub-goal to fail. It
then attempts to find another means for determining the commencement date. This
[30]
is accomplished by attempting to apply an alternative illocutionary operator with

the same goal. An alternative operator it considers is the findings operator whose
argument is notification-by-mail. This unifies successfully with the case data.
As a result, sub-goal (1) noted above is achieved. The sibling sub-goals (2), (3),
and (4) are then achieved in a similar manner. The net result of this process is the
illocutionary structure shown in Figure 9.
The document planner creates the rhetorical structure in a similar way. By
unifying rhetorical operators with the input data, backchaining on these instanti-
ated operators, and recursing, the document planner constructs a fully instantiated
rhetorical structure, which also includes formatting directives. First, the top-most
rhetorical operator posts three sub-goals: (1) construct a "header" which includes
case specific details such as the names of the appellant and appellee and court num-
ber; (2) construct the body of the document; and (3) construct the "footer" which
includes additional case specifics such as the order date. Planning the rhetorical
structure for the body involves emitting the preamble, connective phrases such
as 'furthermore it appears that', and other boilerplate phrases. The net result of
backchaining on rhetorical operators is the rhetorical structure shown in Figure 10.
To enforce consistency of constraints across instantiations of illocutionary and
rhetorical operators, the planner employs a global approach to unification in which
variable bindings in the rhetorical structure are made consistent with bindings in
the illocutionary structure. Illocutionary planning and rhetorical planning are in-
terleaved at runtime. As variables in the operators are bound to the specifics of
the case, the constraints they impose on the instantiation of other operators are
propagated throughout the grammar to create the discourse structure. When docu-
ment planning is complete, the discourse structure contains both the intentional
inferences supporting the findings and also the discourse inferences supporting
the organizational and formatting decisions. The details of the discourse structure
produced by the document planner for Smythe are shown in Appendix B.
The document drafter performs its work in two distinct phases: traversal and
interpretive concatenation. First, it conducts a pre-order traversal of the discourse
structure. For each leaf of the discourse structure, it determines if a text segment
has already been constructed for the node by Inform operators. If none is found, it
must create a text segment that expresses the content in the node. Inform operators
can specify the inclusion of a noun phrase (e.g., 'the trial court's order granting
summary judgment in favor of the plaintiff'), a verb phrase (e.g., 'The notice of
appeal was filed on'), a connective phrase (e.g., 'as to'), a sentence (e.g., 'See
c.A.R. 4(a).'), a formatting directive (e.g., a paragraph break), or a reference to the
facts of a case (e.g., <appellant». To create a text segment, the drafter examines
the node content and constructs a phrase that expresses it. For example, to realize
a date, it extracts the month, day, and year features and transforms them into a date
phrase, (e.g., 'March 27, 1987').
During the interpretive concatenation phase, the document drafter examines
each object produced during the traversal, which includes both text segments and
[31 ]
formatting directives. By concatenating the text segments with the interpreted for-
matting directives in the order specified by the pre-order traversal of the discourse
structure, the drafter produces the final document. 6 For example, the rhetorical
structure created for Smythe includes paragraph breaks and right justifications
that conform to the conventions of show-cause orders issued by the Colorado
Court of Appeals. Figure 4 displays the resulting show-cause order flanked by the
illocutionary and rhetorical structures that produced it.
The DOCU-PLANNER creates documents efficiently. For example, creating the
Smythe show-cause order required 775 milliseconds on a DEC Alpha and Canada
required 937 milliseconds. Typically, document planning per se consumes approx-
imately three quarters of the total execution time while document drafting takes
approximately one quarter of the time.
5. Related Work
Our approach to automated document generation draws on four different lines of
research: discourse structure analysis, the theory of argumentation, explanation
generation, and automated document drafting. The primary focus of research in
discourse structure has been accounting for the coherence of expository or other
communicative text through hierarchical structures of rhetorical and other dis-
course relations, e.g., (Grosz and Sidner, 1986a; Hobbs, 1979). The formalization
of inter-sentential discourse relations is a key requirement for the development of
automated document generation systems.
The most directly relevant portion of research in discourse structure is speech
act theory. Initiated by J.L. Austin, who was primarily concerned with explicit per-
formatives (Austin, 1962), speech act theory addresses the illocutionary content of
discourse, i.e., the goals that a speaker intends to accomplish through that discourse
(Grice, 1975; Searle, 1969).
The theory of argumentation addresses texts intended to persuade, establish,
or prove. For example, Toulmin (Toulmin, 1958) analyzed argumentative texts in
terms of the concepts of warrant, ground, conclusion, backing, and qualification.
This model has been widely applied to the analysis (Marshall, 1989; Zeleznikow
and Stranieri, 1995) and creation (Bench-Capon and Staniford, 1995) of legal doc-
uments. Argument structure, like other forms of illocutionary goal structure but
unlike rhetorical structure, does not directly address the "surface" form of texts.
This line of research is particularly relevant to the analysis of the illocutionary
structure of persuasive or dispositive documents, such as legal briefs and judicial
decisions (B ranting , 1993a).
The explanation community has extensively studied the process of planning and
realizing text given a set of discourse specifications. Over the past decade, their
6 If the user has requested web mode, the formatting directives are first translated to hyper-text
markup language (HTML) formatting commands, which collectively produce a properly formatted
online document.
[32]
research on discourse planning (McKeown, 1985; Paris, 1988; Hovy, 1990; Hovy,
1993; Cawsey, 1992; Suthers, 1993; Moore, 1995; Mittal, 1993; Lester and Porter,
1996) has produced a variety of techniques for determining the content and organi-
zation of many genres of text. Perhaps because of the necessity of coping with the
myriad underlying rhetorical, illocutionary, and argument structures in discourse
generation, this work has yielded a variety of mechanisms for determining the
content and organization of multi-sentential text, a key capability of self-explaining
documents.
Automated document drafting research is the fourth relevant research area.
Two important areas of automated document drafting research are automated le-
gal drafting and automated report generation. A large number of automated legal
drafting systems have been developed in recent years, but most involve creation of
text templates that are then instantiated to create particular documents (Lauritsen,
1992). This approach has been successfully applied to automated drafting of highly
predictable, regular documents (Spirgel-Sinclair, 1988).
Some progress has been made in exploiting explicit representations of the re-
lationship between generic documents and document instances and of constraints
among document components (Daskalopulu and Sergot, 1995). However, there is a
growing recognition in the Law and AI community that a declarative representation
of the knowledge underlying the selection and configuration of textual elements is
essential for the development of tools that embody the expertise of legal drafting
experts (Gordon, 1989; Lauritsen, 1993).
Several recent systems have used declarative representations of legal rules
but only partially declarative representations of rhetorical structure. For exam-
ple, JEDA (Pethe et aI., 1989) used a declarative representation of legal rules,
but mediated document construction entirely through procedural rules. Similarly,
LAW CLERK (Branting, 1993a) used an explicit representation of legal rules and a
simple record structure for administrative law decisions. LAW CLERK instantiated
and wrote text templates associated with predicate/truth-value pairs to the fields of
decision record during back-chaining.
A more detailed declarative rhetorical model was used in PLAID (Bench-Capon
and Staniford, 1995), which produced a document by generating an illocutionary
structure whose nodes were tagged according to their role in the argument (claim,
rebuttal, support, qualification, etc.). This structure was then pruned to exclude
premises which should be implicit in the final presentation and organized into a
structure, including linking text, based on a high-level rhetorical template.
The applied computational linguistics community has addressed the task of
automated report generation from an underlying domain structure. Kittredge et
ai. have observed that representing new domain-dependent discourse knowledge
- they term it "domain communication knowledge" - is required to create ad-
vanced report generators, e.g., for special purpose report planning (Kittredge et aI.,
1991). Given a representation of a particular domain for a particular applica-
tion, knowledge-based report generation is the task of automatically producing
[33 ]
clearly stated reports that are relevant to users of the application. This community
has focused its efforts on deriving technical documentation from program traces
generated during software development or use (Korelsky et aI., 1993; Johnson,
1994; McKeown et aI., 1995) and on producing customized patient information
reports for medical applications (DiMarco et aI., 1995).
6. Discussion and Future Work
In this paper we have presented a model of the illocutionary and rhetorical

structures underlying a representative type of judicial documents - jurisdictional
show-cause orders - and have shown how these structures can be used to form a
document grammar that can generate new documents using a unification-based
procedure. In Section 1 we argued that high standards of correctness and con-
sistency are essential in judicial document drafting, and in Section 2 we argued
for the importance of increased efficiency in drafting routine judicial documents.
We believe that document grammars provide an extremely powerful method for
simultaneously achieving correctness, consistency, and efficiency.
The strength of the discourse structure representation of judicial documents
is that it explicitly represents both an illocutionary structure, which arises from
the applicable domain legal rules, and a rhetorical structure, which expresses the
stylistic and discourse conventions of the genre. Earlier approaches to routine ju-
dicial document drafting generally omitted one aspect of the discourse structure or
conftated the two.
We advocate an approach that uses explicit representations of illocutionary and
rhetorical structures in a document grammar that applies a unification-based for-
malism to this grammar to create documents. There are two key advantages to
this approach. First it aids knowledge acquisition and system building because it
facilitates a conceptual model that cleanly separates (1) substantive domain rules,
e.g., the requirements for jurisdiction, (2) genre-specific stylistic and discourse
conventions, and (3) the procedure for creating documents that conform to the
domain rules and genre conventions. Illocutionary operators can be conceptualized
in terms of the underlying legal rules, without consideration of the way in which
the rules will be embedded in a document. Similarly, rhetorical operators can be
formalized independently of the legal rules that the document will ultimately ex-
press. Both forms of operators can be developed without requiring consideration
of the procedures governing their use in document drafting.
A second, related, advantage is that this approach assists validation (because
each component can be tested separately) and maintenance (because either the
illocutionary or the rhetorical operator set can be modified without affecting the
other). For example, a change in jurisdictional rules can, in general, be accommo-
dated by modifying the illocutionary operators without requiring any changes in
the rhetorical operators.
[34]
This paper has focused on document drafting ab initio given a relatively com-
plete document grammar. We believe that the dual-justification structure also has
important potential benefits for retrieval, explanation, and adaptation of existing
documents, comparison of alternative drafts of documents at a "deep" illocutionary
level, and maintenance of multi-generation documents. Document drafting through
reuse is a promising method of extending the coverage of a document grammar,
since retrieval and adaptation can be performed with even an incomplete set of illo-
cutionary operators. Accordingly, the first step in our research agenda is developing
mechanisms that use discourse structure for retrieval, explanation, comparison, and
interactive adaptation of existing documents.
This paper has illustrated the creation of new documents from a document gram-
mar that was manually constructed. However, constructing a document grammar is
a challenging activity that currently requires considerable knowledge of the un-
derlying unification formalism. Therefore, the second step in our current research
agenda is to develop a semi-automated knowledge acquisition tool for interactively
acquiring document grammars. We envision that domain experts, beginning with a
core document grammar, will use this tool to construct document grammars that ex-
hibit increasingly greater coverage. However, rather than expressing the document
grammar directly in the syntactically complex unification formalism, the grammar
acquisition tool will enable them to state illocutionary and rhetorical operators in
a high-level language and help them rapidly evaluate the coverage of the grammar
until it generates precisely the set of documents they intend.
7. Summary
This paper has presented a model of document structure that makes explicit (1)
the document's illocutionary structure, i.e., the connection between the document
drafter's goals and the text intended to achieve those goals, and (2) the document's
rhetorical structure i.e., the stylistic and discourse conventions of the documen-
t's genre. This model was applied to a representative class of judicial orders,
jurisdictional show-cause orders. The ability of a document grammar based on
the illocutionary and rhetorical structure of representative documents to synthe-
size additional documents was then illustrated with a simple example. Practical
knowledge-based judicial document drafting systems for routine judicial docu-
ments would make a significant contribution to judicial efficiency. The approach
described in this paper is a first step towards the development of such practical
systems.
[35 ]
Appendix A
THE COMPLETE DOCUMENT GRAMMAR
This appendix presents the document grammar used to synthesize the show cause
orders for Smythe, Kirkpatrick, and Canada. For clarity, the grammar is ex-
pressed in a context-free-style syntax; it is in fact implemented in the functional
unification language, FUF, whose syntax is significantly more complex. Expres-
sions on the left-hand side of each production are predicate non-terminals, e.g.,
Rule(Untimely-Notice-Of-Appeal), while the right-hand side consists of one or more
combinations of (1) a predicate non-terminal, (2) a collective non-terminal e.g.,
At-Least-One[X, Y, Z], (3) a conditionalized non-terminal, e.g., If{condition}, and
(4) a variable, which is indicated with braces, e.g., <var>. The null expression
is indicated with E. Ordering constraints, which are represented with Sequence
predicates, are after the productions.
Establish(Preqs-For-Dismissal) ==> Establish(Jurisdictional-Defect)

Order(Show-Cause)
Establish(JurisdictionaI-Defect) ==> Establish(Appeals)

At-least-One [ Establish(Untimely-Notice-Of-Appeal),
Establish(Subject-Defect),
Establish(Finality-Defect) 1
Establish(Appeals) ==> One-Of [ Find(No-Main-Appeal),

Find(No-Follow-Up-Appeal),
Find(Follow-Up-Appeal) 1
Establish(Untimely-Defect) =~ IF <judgment-date>
{ Establish(Untimely-NOA) }
Establish(Subject-Defect) ==> IF <subject-matter>

{ Establish(Subject-Matter-Defect) }
Establish(Untimely-NOA) ==> IF <judgement-date>

{ Establish(NOA-Commencement-Date)
Establish(NOA-Filing-Date)
Establish(NOA-Due-Date)
Rule(Untimely-NOA) }
Establish(NOA-Cornmencement-Date) ==> Find(Judgment)

Find(NOA-Commencement)
Establish(NOA-Filing-Date) :~ Find(NOA-Filing-Date)
Establish(NOA-Due-Date) =0> Find(NOA-Due-Date)
[36]
Establish(Finality-Defect) =? IF <order-appealed-from>
{ Establish(Order-Appealed-From)
Establish(Non-Final-Order) }
Establish(Non-FinaI-Order) =? Rule(Non-FinaI-Order)
Establish(Order-Appealed-From) =? Find(Appeal-Source)
<order-appealed-from>
Inform(Appeal-Due) =? "the notice of appeal was due"
Inform(One-AppeaI-Only) =? "judgment was entered"
Inform(Main-AppeaI-Entered) =? "was entered"
Inform(Appeal-Source) =? "this appeal is from"
Inform(Untimely-Appeal) =? "the notice of appeal was not timely filed."
Inform(Untimely-Main-Appeal) =? "the notice of appeal was not timely"
Inform(Untimely-NOA-Rationale) =? "for failure to file a timely notice of appeal."
Inform(Appeal-Filed) =? "the notice of appeal was filed on"
Inform(FinaI-Sanction) =? "why this appeal should not be dismissed with prejudice"
Inform(Non-Final-Sanction) =? "why this appeal should not be dismissed without prejudice"
Inform(Partial-FinaI-Sanction) =? "why this appeal should not be partially dismissed with

prejudice to the extent that defendant seeks review of"
Inform(PartiaI-Non-Final-Sanction) =? "wby this appeal should not be partially dismissed

witbout prejudice to the extent that defendant seeks
review of"
Inform(Non-Final-Judgment) =? "this is not a final judgment because it does not end

'the particular action in which it is entered, leaving
nothing further for the court pronouncing it to do in
order to completely determine the rights of the parties
involved in the proceeding.' "
Inform(Appeal) =? "defendant is appealing from"
Inform(Appeal-Both) =? "defendant is appealing from both"
Inform(Non-Final-Order-Rationale) =? "for failure to file a final appealable order."
Inform(M2iled) =? "and mailed to counsel of record on"
Order(Show-Cause) =? Order(Response-Due-Date)
One-Of [Order(Sanction),
Order(PartiaI-Sanction) 1
One-Of [Order(Non-Final-Order-Rationale),
Order(Untimely-NOA-Rationale) 1
Order(Sanction) =? IF NOT <main-appeal>

{ One-Of [ Order(Non-Final-Sanction),
Order(Final-Sanction) 1}
[37 ]
Order(Partial-Sanction) ~ IF <main-appeal>
{ One-of [ Order(Partial-Non-FinaI-Sanction),
Order(Partial-Final-Sanction) 1
<main-appeal> }
Order(Non-FinaI-Order-Rationale) ~ IF <order-appealed-from>
{ Inform(Non-Final-Order-Rationale) }
Order(Untimely-NOA-Rationale) ~ IF NOT <order-appealed-from>

{ Inform(Untimely-NOA-Rationale) }
Order(Response-Due-Date) ==? Compute-Response-Date( <judgment-date>, <mailing-date»
Order(Non-FinaI-Sanction) =* IF <order-appealed-from>
{ Inform(Non-FinaI-Sanction) }
Order(Final-Sanction) ~ IF NOT <order-appealed-from>

{ Inform(FinaI-Sanction) }
Order(Partial-Non-FinaI-Sanction) ~ IF <order-appealed-from>
{ Inform(Partial-Non-FinaI-Sanction) }
Order(Partial-Final-Sanction) ~ IF NOT <order-appealed-from>

{ Inform(PartiaI-FinaI-Sanction) }
Find(Judgment) ~ One-Of [ Find(One-AppeaI-Only),

Find(Main-Appeal) 1
Find(One-Appeal-Only) ~ IF NOT <main-appeal>

{ Inform(One-AppeaI-Only) }
Find(Main-Appeal) ~ IF <main-appeal>
{ Inform(Main-AppeaI-Entered) }
Find(NOA-Commencement) '=- One-Of [ Find(Mailed),

Find(Present) 1
Find(Mailed) ~ IF EQU AL( <notification-method>, mailing)
{ <judgment-date>
Inform(Mailed)
<mailing-date> }
Find(Present) ~ IF EQUAL«notification-method>, present-at-judgment)

{ <judgment-date> }
Find(No-Main-Appeal) ~ IF NOT <main-appeal>

{ Inform(Appeal-Source)}
Find(No-Follow-Up-Appeal) '= IF NOT <follow-up-appeal>

{ Inform( Appeal)
<main-appeal> }
Find(Followup-Appeal) ~ IF <followup-appeal>
{ Inform( Appeal-Both)
<main-appeal>
<follow-up-appeal> }
Find(NOA-Due-Date) ~ Inform(Appeal-Due)
Compute-Due-Date( <judgment-date>, <mailing-date»
Find(AppeaI-Source) ~ Inform(AppeaI-Source)
[38]
Find(NOA-Filing-Date) ~ Link(And)
Inform(Appeal-Filed)
<filing-date>
Rule(Untimely-NOA) ~ One-Of [ Rule(No-Main-Appeal),

Rule(Main-Appeal) 1
Rule(Authority)
Rule(No-Main-Appeal) ~ IF NOT <main-appeal>

{ Rule(Untimely-Appeal) }
Rule(Main-Appeal) ~ IF <main-appeal>
{ Rule(Untimely-Main-Appeal) }
Rule(Untimely-Main-Appeal) ~ Inform(Untimely-Main-Appeal)
<main-appeal>
Rule(Untimely-Appeal) ~ Inform(Untimely-Appeal)
Rule(Non-Final-Order) ~ Inform(Non-FinaI-Judgment)
Rule(Authority) ~ <authority>
Frame(Header) ~ "Colorado Court of Appeals Order"

<case-number>
<order-number>
Frame(Referent)
<appellant>
" Appellant"
"and"
<appellee>
" Appellee"
"To: "
<appellant>
Frame(Attorneysl
Frame(Referent) ~ One-Of [ Frame(Domestic),

Frame(Civil) 1
Frame(Domestic) ~ IF EQUAL( <substantive-legal-area>, domestic)

{ "In re tbe Marriage of" }
Frame(Civil) ~ IF EQUAL( <substantive-legal-area>, civil)

{ <attorneys> }
Frame(Attorneys) ~ One-Of [ Frame(Not-Pro-Se),

Frame(Pro-Se) 1
Frame(Not-Pro-Se) ~ IF <attorneys>
{ <attorneys> }
Frame(Pro-Se) ~ "pro se appellant"
Frame(Body) ~ Link(Preamble)
Link(However)
Link(Furthermore)
Link(Thus)
Link(Show-Cause)
Frame(Body) ~ Link(Preamble)
Link(Further)
Link(Show-Cause )
[39]
Frame(Body) ==> Link(Preamble)

Link(Furthermore)
Frame(Footer) ==> " BY THE COURT"

"Date: "
<order-date>
"Copies to: Counsel of Record"
Link(Preamble) ==> "From the notice of appeal filed by appellant and the
district court, it appears that"
Link(Thus) ==> "Thus, it appears that"

Link(Further) ==> "It further appears that"
Link(Furthermore) ==> "Furthermore, it appears that"
Link(Show-Cause) ==> "IT IS THEREFORE ORDERED that the appellant shall
show cause, if any there is, in writing on or before"
link(However) ==> "However, it appears that"

Link(As-To) ==> "as to"
Link(And) ==> " and"
Organize(Show-Cause-Order) ==> Frame(Header)
Frame(Body)
Frame(Footer)
Sequence(Preqs-For-Dismissal) = left-To-Right Frame(Header),

Link(Preamble),
Establish( J urisd ictional- Defects),
Order(Show-Cause ),
Frame(Footer) J
Sequence(Untimely-NOA) = Left-To-Right [ Establish(NOA-Commencement),

Establish(NOA-Filing-Date),
link(Furthermore),
Establish(NOA-Due-Date),
link(Thus),
Rule(Untimely-NOA) J
Sequence(Finality-Defect) = Left-To-Right [ Establish(Order-Appealed-From),

Link(Further),
Establish(Non-FinaI-Order) J
Sequence(Show-Cause) = left-To-Right Link(Response-Due-Date),

Order(Response-Due-Date),
Order(Sanction) I Order(Partial-Sanction),
Order(Non-FinaI-Order-Rationale) I
Order( Unti mely-N OA-Rationale)
Sequence(Appeals) = left-To-Right [ Inform(Appeal),

<main-appeal>,
Link(However) J
Sequence(Appeals) = left-To-Right [ Inform(Appeal-Both),

<main-appeal>,
link(And),
<followup-appeal> ,
link(However) 1
Sequence(Untimely-Main-Appeal) = left-To-Right [ Inform(Untimely-Main-Appeal),

Link(As-To),
<main-appeal> 1
[40]
AppendixB
THE DISCOURSE STRUCTURE CONSTRUCTEDF OR SMYTHE
This appendix presents the discourse structure produced by the document plan-
ner by unifying the Smythe case facts with the document grammar for generating
show-cause orders. The instantiated illocutionary and rhetorical structures are in-
terleaved via the sequencing constraints. For purposes of presentation, the structure
is depicted as a strict hierarchy rooted at (Establish(Preqs-for-Dismissal)). The doc-
ument drafter interprets this structure to create the Smythe show-cause order shown
in Figure 4.
EstabliBh(PreqB-for-DismiBsal) -->
Frame (Header) -->
"Colorado Court of Appeals Order"
<case-number> • "No. 96CA0437"
<order-number> • "Tr. Ct. No. 91CV061"
Frame(Referent) -->
Frame (DomeBtic) -->
"In re the Marriage of"
<appellant> • "Herbert W. Smythe"
Appellant"
II and"
<appellee> • "Catherine Smythe"

Appellee"
liTo: It
<appellant> • "Herbert W. Smythe"

Frame(AttorneYB) -->
Frame(Pro-Se) -->
"pro Be appellant"
LiDk(Preamble) -->
"From the notice of appeal filed by appellant and the
district court, it appears that"
EstabliBh(JuriBdictional-Defect) -->
EBtabliah(Appeals) -->
Find(No-Main-Appeal) --> e
Establish(Untimely-Defect) -->
Estab11sh(Untimely-NOA) -->
EatabliBh(NOA-Commencement-Date) -->
Find (Judsment) -->
Find(One-Appeal-Only) -->
[ 41 ]
Inform(One-Appeal-Only) -->
"judgment vas entered"
Find(NOA-Commencement) -->
Find(Kailed) -->
<judgment-date> • "September 20, 1996"
Inform(Kailed) -->
"and mailed to counsel of record on"
<mailing-date> • "September 22, 1996"
Establish(NOA-Filing-Date) -->
Find(NOA-Filing-Date) -->
Link(And) -->
"and"
Inform(Appeal-Filed) -->
"the notice of appeal vas filed on"
<filing-date> • "November 7, 1996"
Establish(NOA-Due-Date) -->
Find(NOA-Due-Date) -->
Link(Furthermore) -->
"Furthermore, it appears that"
Inform(Appeal-Due) -->
"the notice of appeal vas due"
Compute-Due-Date«judgment-date>, <mailing-date»
• "November 6, 1996"
Rule (Untimely-NOA) -->
Rule (No-Kain-Appeal) -->
Link(Thus) -->
"Thus, it appears that"
Inform(Untimely-Appeal) -->
"the notice of appeal vas not timely filed"
Rule(Authority) -->
<authority> • "See C.A.R. 4(a)."
Order(Shov-Cause) -->
Link(Shov-Cause) -->
"IT IS THEREFORE ORDERED that the appellant shall
shov cause, if any there is, in vriting on or before"
Order (Response-Due-Date) -->
Compute-Response-Date«judgment-date>, <mailing-date»
• "December 18, 1996"
Order (Sanct ion) -->
Order (Final-Sanction) -->
Inform(Final-Sanction) -->
"'vhy this appeal should not be dismissed vith prejudice"
Order(Untimely-NOA-Rationale) -->
Inform(Untimely-HOA-Rationale) -->
"for failure to file a timely notice of appeal"
Frsme(Footer) -->
BY THE COURT"
"Date: "
<order-date> • "December 4, 1996"
"Copies to: Counsel of Record"
Acknowledgments
The authors wish to thank: Anne Gill, staff attorney of the Colorado Court of
Appeals, for assisting us in understanding jurisdictional screening procedures;
Michael Elhadad of Ben Gurion University for his generous assistance with FUF;
[42 ]
Bruce Porter of the University of Texas at Austin for sharing the computing re-
sources of his laboratory; Trevor Bench-Capon of the University of Liverpool,
Giovanni Sartor of The Queen's University of Belfast, and Gary Stelling and Stu-
art Towns of North Carolina State University for insightful comments on earlier
drafts of this manuscript; and Stuart Towns for assistance in creating the figures.
This research is supported in part by a grant from the National Center for Auto-
mated Information Research and by NSF Faculty Early Career Development Grant
IRI-9502152.
References
Allen, J. (1987) Natural Language Understanding. Benjamin/Cummings, Menlo Park, California.
Appelt, D .E. (1985) 'Planning English referring expressions' , Artificial Intelligence 26, 1-33.
Austin, J. (1962) How to Do Things with Words. Oxford U. Press, New York.
Bench-Capon, T. and Staniford, G. (1995) 'PLAID - proactive legal assistance', In Proceedings of
the Fifth International Conference on Artificial Intelligence and Law, pp. 81-88.
Branting, L. and Lester, J.e. (1996a) 'A framework for self-explaining legal documents', In Pro-
ceedings of the Ninth International Conference on Legal Knowledge-Based Systems (JURIX-96),
Tilburg University, the Netherlands, pp. 77-90.
Branting, L. and Lester, J.e. (1996b) 'Justification structures for document reuse', In Proceedings
of the Third European Workshop on Case-Based Reasoning (EWCR-96), Lausanne, Switzerland,
pp.76-90.
Branting, L.K. (1993a) 'An issue-oriented approach to judicial document assembly', In Proceed-
ings of the Fourth International Conference on Artificial Intelligence and Law, Amsterdam, The
Netherlands. ACM Press, pp. 228-235.
Branting, L.K. (1993b) 'A reduction-graph model of ratio decidendi', In Proceedings of the Fourth
International Conference on Artificial Intelligence and Law, Amsterdam, The Netherlands. ACM
Press, pp. 40-49.
Cawsey, A. (1992) Explanation and Interaction: The Computer Generation of Explanatory Dia-
logues. MIT Press.
Cohen, P.R. and Perrault, e.R. (1979) 'Elements of a plan-based theory of speech acts', Cognitive
Science 3,177-212.
Daskalopulu, A. and Sergot, M. (1995) 'A constraint-driven system for contract assembly', In
Proceedings of the Fifth International Conference on Artificial Intelligence and Law, pp. 62-70.
DiMarco, C., Hirst, G., and Wanner, L. (1995) 'HealthDoc: Customizing patient information and
health education by medical condition and personal characteristics', In Working Notes of the
Workshop on Artificial Intelligence in Patient Education.
Elhadad, M. (1991) 'FUF: The universal unifier user manual version 5.0. Technical Report CUCS-
038-91', Department of Computer Science, Columbia University.
Elhadad, M. (1992) Using Argumentation to Control Lexical Choice: A Functional Unification
Implementation. PhD thesis, Columbia University.
Gordon, T. (1989) 'A theory construction approach to legal document assembly', In Pre-Proceedings
of the Third International Conference on Logic, Informatics, and Law, Florence, pp. 485-498.
Grice, H. (1975) 'Logic and conversation', In Cole, P. and Morgan, J. (eds.), Syntax and Semantics
2: Speech Acts, Academic Press, New York, N.Y., pp. 41-58.
Grosz, B. and Sidner, C. (l986a) 'Attention, intention, and the structure of discourse', Computational
Linguistics 12(3).
Grosz, B.J. and Sidner, e.L. (1986b) 'Attention, intentions, and the structure of discourse',
Computational Linguistics 12(3): 175-204.
[43 ]
Hobbs, J. (1979) 'Coherence and co-reference', Cognitive Science 3(1), 67-82.

Hovy, E.H. (1990) 'Pragmatics and natural language generation', Artificial Intelligence 43,153-197.
Hovy, E.H. (1993) 'Automated discourse generation using discourse structure relations', Artificial
Intelligence 63,341-385.
Johnson, W.L. (1994) 'Dynamic (re)generation of software documentation', In Proceedings of the
Fourth Systems Reengineering Technology Workshop, pp. 57-66.
Kay, M. (1979) 'Functional grammar', In Proceedings of the Berkeley Linguistic Society.
Kittredge, R., Korelsky, T., and Rambow, O. (1991) 'On the need for domain communication
knowledge', Computational Intelligence 7(4), 305-314.
Korelsky, T., McCullough, D., and Rambow, O. (1993) 'Knowledge requirements for the automatic
generation of porject management reports', In Proceedings of the Eigth Knowledge-Engineering
Conference. IEEE Computer Society Press.
Lauritsen, M. (1992) 'Technology report: Building legal practice systems with today's commericial
authoring tools', Law and Artificial Intelligence 1(1).
Lauritsen, M. (1993) 'Knowing documents', In Fourth International Conference on Artificial
Intelligence and Law" Amsterdam. ACM Press, pp. 185-19l.
Lester, J.e. and Porter, B.w. (1996) 'Scaling up explanation generation: Large-scale knowledge
bases and empirical studies', In Proceedings of the Thirteenth National Conference on Artificial
Intelligence, pp. 416-423.
Lester, J.e. and Porter, B.W. (1997) 'Developing and empirically evaluating robust explanation
generators: The KNIGHT experiments', Computational Linguistics 23(1), 65-1Ol.
Mann, W.e. and Thompson, S.A. (1987) 'Rhetorical structure theory: A theory of text organization',
Technical Report ISURS-87-190, USClInformation Sciences Institute, Marina del Rey, CA.
Marshall, C. (1989) 'Representing the structure of a legal argument', In Proceedings of the Second
International Conference on Artificial Intelligence and Law, Vancouver, B.C., pp. 121-127.
Maybury, M.T. (1992) 'Communicative acts for explanation generation', International Journal of
Man-Machine Studiej 37(2), 135-172.
McKeown, K. (1982) Generating Natural Language Text in Response to Questions about Database
Structure. PhD thesis, University of Pennsylvania.
McKeown, K. (1985) Text Generation: Using Discourse Strategies and Focus Constraints to
Generate Natural Language Text. Cambridge University Press.
McKeown, K., Robin, J., and Kukick, K. (1995) 'Generating concise natural language summaries',
Information Processing and Management. Special Issue on Summarization.
Mittal, V. (1993) Generating Natural Language Descriptions with Integrated Text and Examples.
PhD thesis, University of Southern California.
Moore, J.D. (1995) Participating in Explanatory Dialogues. MIT Press.
Moore, J.D. and Paris, CL (1993) 'Planning text for advisory dialogues: Capturing intentional and
rhetorical information', Computational Linguistics 19(4),651-694.
Moore, J.D. and Swartout, W.R. (1991) 'A reactive approach to explanation: Taking the user's
feedback into account', In Paris, C.L., Swartout, W.R., and Mann, W.C. (eds.), Natural Lan-
guage Generation in Artificial Intelligence and Computational Linguistics, Kluwer Academic
Publishers, Boston, pp. 3-48.
Paris, C. (1988) 'Tailoring object descriptions to a user's level of expertise', Computational
Linguistics 14(3),64--78.
Pethe, V.P., Rippey, e.P., and Kale, L.v. (1989) 'A specialized expert system for judicial decision
support', In Proceedings of the Second International Conference on Artificial Intelligence and
Law, Vancouver, B.e., pp. 190-194.
Searle, J. (1969) Speech Acts: An Essay in the Philosophy of Language. Cambridge University Press,
Cambridge.
Snellenburg, A.e. (1989) 'New approaches to reducing court delay and congestion', State Court
Journa/13(3).
[44]
Spirgel-Sinclair, S. (1988) 'The DHSS retirement pension forecase and advice system', In Duffin, P.
(ed.), KBS in Government 88, Blenheim On Line, Pinner, pp. 89-106.
Suthers, D.D. (1991) 'A task-appropriate hybrid architecture for explanation', Computational
Intelligence 7(4), 315-333.
Suthers, D.D. (1993) An Analysis of Explanation and Its Implicationsfor the Design of Explanation
Planners. PhD thesis, University of Massachusetts.
Toulmin, S.E. (1958) The Uses of Argument. Cambridge University Press.
Zeleznikow, J. and Stranieri, A. (1995) 'The Split-Up system: Integrating neural networks and rule-
based reasoning in the legal domain', In Proceedings of the Fifth International Conference on
Artificial Intelligence and Law, pp. 185-194.
[45 ]
Artificial Intelligence and Law 6: 151-202, 1998.
151
Criminal Sentencing and Intelligent Decision

Support
URI J. SCHILD
Department of Mathematics and Computer Science, Bar-Ilan University, Ramat-Gan 52900, Israel
Abstract. The construction of computer systems for sentencing decision support has lately been
of increasing interest. Different approaches and models exist for such systems. Of the available
technologies we advocate improved access to precise statistical data, reject a rule-based expert system
approach, and argue that a case-based advisory system approach could be the most acceptable. Two
prototype systems we have under development and testing are described.
"There are two kinds of people in prison: those who should

never have been sent there, and those who should never be let
out."
(George Bernard Shaw)
1. Introduction
1.1. THE PROBLEM
The problem described here deals with the question how computer systems may
assist judges in the process of passing sentence in criminal cases. As should seem
obvious the sentencing process calls for human discretion: Apart from mandatory
sentences for specific offences, judges are usually faced with having to choose one
out of many acceptable sentences. This process of selection and decision malting is
a most distinctive human task, and attempts at its computer simulation would form
one of the major challenges for Artificial Intelligence.
The present state-of-art in AI does not enable such a simulation. As a matter of
fact, most people - computer scientists, legal experts and laymen - do not want to
consider the development of computer systems that independently would carry out
the tasks of a judge. They believe that as a matter of principle this activity should
be an exclusively human one (see (Gardner 1987, 79-83) for a fuller discussion of
this issue, and other problems relating to misuse of computer programs in law).
The goal should rather be to build a decision support system for sentencing.
Such a system would supply a judge with relevant information and advice. The
operation of the system could be based on statistical or criminological models, and
[47 ]
152 URI J. SCHILD
it could also involve an expert knowledge-base. Information would be retrieved by

the system through an intelligent reasoning process. But the final decision would
be made by the judge himself. 1
The construction of such a system is also a major challenge. It is our goal to
discuss what kind of support a computer system can give to a judge with respect to
the sentencing process. We shall consider what other researchers have done on this
subject, and also describe our own work. We shall make a distinction between non-
intelligent computer systems, which are essentially statistical or criminological in
nature, and expert systems, i.e., systems which exhibit intelligence of their own.
The concept of an expert system does not have an exact definition anymore (if
it ever had one). We5hall call the kind of system that passes sentence on its own,
an expert system, while the kind of system we are interested in could be called an
intelligent advisory system.
The discussion of our problem is also relevant to similar problems in other
areas. Determination of amounts of child support and alimony, or calculation of
reparations for physical injuries are problems where human discretion is applied
to fixing a numerical value. There may also be similar problems outside the legal
domain.
1.2. PRACTICAL APPLICATION

The topic of legal computer systems for sentencing is of interest both from a the-
oretical jurisprudential point of view, and a criminological one. It is of course of
relevance for computer scientists working in the area of AI and Law, and it can also
have a major practical importance, as we shall now explain.
The main reason for developing and applying sentencing advisory systems re-
lates to the quality of the sentencing process as viewed by the public at large and
by the judiciary itself. Consider the following story by John Dewey as quoted in
(Gardner 1987, 83):
There is the old story of the layman who was appointed to a position in India
where he would have to pass in his official capacity on various matters in
controversy between the natives. Upon consulting a legal friend, he was told
to use his common-sense and announce his decisions firmly; in the majority
of cases his natural decision as to what was fair and reasonable would suffice.
But his friend added: Never try to give reasons, for they will usually be wrong.
(Dewey 1924, 17)
This story is perhaps amusing, but it mainly illustrates some fundamental differ-
ences between non-professional and professional adjudicators. A sense of what is
fair and reasonable is of course a quality one would expect in any kind of adjudica-
tor. In past days of imperialism and even today in the lowest level of the courts this
1 In the following 'he', 'him', 'himself', etc. should not be understood as gender-specific, but also
as including 'she', 'her', 'herself', etc.
[48]
CRIMINAL SENTENCING AND INTELLIGENT DECISION SUPPORT 153
may perhaps be sufficient qualification for a lay-judge. However, a professional

judge has a far greater responsibility He must always state the reasons for his
decisions, and both the decisions and their justification can be directly relevant
and have impact on later decisions of other judges.
Concepts like 'fair and reasonable', not to mention 'just and equitable', are not
always obvious for a case at hand, and even a professional judge may arrive at a
wrong decision or give wrong reasons for a correct decision, though the legal sys-
tems of the world usually have options for appeal, etc. When the decision of a judge
has to be quantitative, it is even more difficult to evaluate and compare to other
decisions, and it sometimes happens that the public expresses its dissatisfaction
with the quality of the sentencing of this judge or another.
In the domain of criminal sentencing it appears that even when giving more
or less identical reasons, judges arrive at very different conclusions. In practice
it appears that in many instances there is a great disparity in the decisions of
sentencing judges - even in quite similar cases. It also happens that the very same
judge will decide on vastly different sentences in similar cases occurring even over
a relatively short time-span.
It is important to understand why this happens, and why some of those differ-
ent conclusions cannot just be discounted as being erroneous. One of the reasons
relates to different sentencing approaches. When passing sentence, a judge will
usually take one or more of the following fundamental sentencing approaches into
account: Retribution, Deterrence, Prevention and Rehabilitation (see Sections 2.1
and 2.3). Different judges may decide sentences in similar cases from different
viewpoints and thus reach different results.
Sentencing policy is subject to changes. Usually such changes are initiated by
legislators and the highest levels of the judiciary. But individual judges may also
change their own, private sentencing policies with time. A judge we interviewed
told us, that when he took up his position on the Bench, he often decided sentences
with a view to rehabilitate the offender. "Now", he said, "I believe that the profes-
sional criminal cannot be rehabilitated, and it is only a matter of keeping him out
of circulation". 2
It is universally assumed that a judge passing sentence in a given case should
follow trends established in sentencing of similar cases. There is therefore often
some dissatisfaction with the sentencing process as it takes place today in the
courts of many countries. A computer system for sentencing support could possibly
enable the judiciary to pass sentences of greater uniformity, without, of course,
impairing their freedom and independence.
2 This private opinion also happens to reflect the general trend in many countries during the last
10-20 years. See also Lawton L.I. in Davies (1978) Cr. App. Rep. 207 C.A.
[49 ]
154 URI J. SCHILD
1.3. INTERACTION WITH ISRAELI JUDICIARY

We have been very fortunate that the Israeli judiciary at all levels has supported us
in this research. Our work has not only been viewed with interest, but with active
cooperation. The Israeli judicature has three levels, and several judges from the Tel
Aviv Districts and Appeals Court (the intermediate level) have supplied us with
expert knowledge about sentencing and tried our prototype systems.
Their main motivation is not their great interest in Computer Science in general
or Artificial Intelligence in particular, but relates to their vehement opposition to
the introduction of sentencing guidelines (see below, Section 3.1) in Israel. Some
believe that any computer system we may come up with will be better than sen-
tencing guidelines. Others fear that the politicians will attempt to make sentencing
guidelines mandatory, but hope that the availability of a computer support system
(any kind of system) will put an end to that attempt. The use of a computer support
system would presumably always be optional and voluntary, so its introduction is
not viewed as a problem.
1.4. CONTENTS OF PAPER

The outline of this paper is as follows. In section two we shall give some back-
ground relating to the criminology of sentencing. In section three we shall consider
work that relates to non-intelligent computer systems for sentencing support. Sec-
tion four will deal with intelligent sentencing systems, mainly rule-based. The next
two sections describe the case-based approach and uses a system we have devel-
oped as an example. Section seven analyses and compares the various approaches
with respect to maintenance, scale-up and efficiency.
2. Criminology of Sentencing
A sentence determines some form of punishment. It is therefore natural that we
consider the the criminological aspects of punishment, especially as uniformity of
sentencing (and hence of punishment) is a central issue in our work.
2.1. FRAMEWORKS FOR PUNISHMENT
Punishment and its purposes may be examined from three different aspects. The
first one is the philosophy of punishment, the second is the framework of law and
order in society, and the third is often called the sociology of punishment (Garland
1991).
The philosophy of punishments deals with questions like: What is the justi-
fication for inflicting punishment? What is a just punishment? etc. The second
approach (law and order) deals with fighting crime. Within this framework criminal
punishment has one specific aim: Reduction and prevention of crime. Punishment
[50]
is here considered, not from the viewpoint of justice, but from the viewpoint of ef-
ficiency. Which punishments will deter criminals? Can criminals be rehabilitated?
etc. The third framework considers society and punishment, and the connection
between them.
The four classical approaches to punishment, Retribution, Deterrence, Preven-
tion and Rehabilitation form another type of classification of punishment com-
monly used by the judiciary and by criminologists:
We have thought it necessary not only to analyse the facts, but to apply to
those facts the classical principles of sentencing. Those classical principles
are summed up in four words: retribution, deterrence, prevention and rehabil-
itation. Any Judge who comes to sentence ought always to have those four
classical principles in mind and to apply them to the facts of the case to see
which of them has the greatest importance in the case with which he is dealing
(Lawton L.J., in: Sargeant (1974) 60 Cr. App. Rep. 74 c.A. at pp. 77-84).
These four sentencing aims are not easily incorporated within the first men-
tioned frameworks, as the concepts may conflict with each other: A retributive
punishment may be just but not efficient, and vice versa. Some philosophers will
argue that deterrence is not just. Why should one person be punished in order to
deter others from committing crimes?
It should therefore be clear that different persons may hold very different views
relating to the concept of punishments, and that society as such may change those
views over time. As an example of totally conflicting views on punishment consider
the following quotes from two eminent legal authorities.
The punishment for grave crimes should adequately reflect the revulsion felt
by the great majority of citizens for them: The ultimate justification of any
punishment is not that it is a deterrent, but that it is the emphatic denunciation
by the community of a crime (Denning 1953, para. 53)
The idea that we may punish offenders against a moral code, not to prevent
harm or suffering or even repetition of the offence but simply as a means
of venting or emphatically expressing moral condemnation, is uncomfortably
close to human sacrifice as an expression of religious worship (Hart 1963,
65-66)
It follows that when examining any concept relating to punishment, one should
consider both the appropriate framework and the underlying criminological ap-
proach. Thus, the motivation for a given sentence in a particular case must be
analysed with respect to these concepts, and the sentence cannot not be compared
to the sentence of another - superficially similar - case without further inquiry.
This now leads us to consider the problem of sentencing disparity.
[51]
156 URI J. SCHILD
2.2. UNIFORMITY OF PUNISHMENT

It seems obvious that the request for uniformity in punishment is more closely
related to the search for justice than to prevention of future crime. The reaction to
a serious crime ought to be stronger than to a petty crime, and people convicted of
the same crime ought to receive the same punishment. One of the leading crimi-
nologists, Andrew VOll Hirsch, who is active in sentencing reform, has expressed
this as follows:
A sentencing system should seek to be just - or at least, to be as little unjust
as possible. Claims about fairness, not. crime-control effectiveness, underlie
the requirements of proportionality. (von Hirsch 1993, 103)
However, within a crime-fighting framework uniformity may be viewed as an
obstacle. Should we wish to rehabilitate or deter offenders, then the punishment
should depend more on the particular offender than on the particular crime he
has committed. In many cases it is clear, that a prison sentence will only make
an offender return to crime. In other cases it is equally clear, that only a prison
sentence will prevent an offender from committing another crime, even if he has
been found guilty of a minor crime. On the other hand, it has been argued that
a non-uniform sentencing system can bring about feelings of frustration among
prisoners, leading to unrest in the prisons.
The public at large expects some kind of fairness expressed through uniform
sentences. However, the public also wants law and order, even though uniformity
and effectiveness of punishment to some extent oppose each other. In reality the
judges must seek some kind of equilibrium among the various approaches.
Such equilibrium is not easy to attain, and both politicians and the media see
themselves involved. Israel is not the only country where the present judicial sen-
tencing process is being criticised. In England, for example, the Home Secretary
has strongly advocated mandatory minimum sentences for several crimes (Home
Office White Paper, 1996), with equally strong opposition from the judiciary and
other parties involved. Such a proposal is of course popular with a public that
believes that its adoption will contribute to law and order. 3
We shall give yet another example of the problems and the bias involved in
finding the appropriate equilibrium. There is in Israel (and probably in many other
countries) a general belief - strongly supported by specific examples in the media
- that rape sentences are far too lenient. Information on offenders and sentences
are stored in the central police files. No unauthorised personnel has access to this
information, and authorisation is only given on an extremely limited basis. From
our work in the sentencing domain we know for a fact that no statistical study has
ever been made on the severity of rape sentences (or the lack thereof). Yet, the
media make their assertions again and again based on a small number of cases they
select for that particular purpose.
3 The Home Office Recommendations have indeed been adopted by Parliament in the Crimes
(Sentencing) Act 1997.
[52]
We must therefore amplify the views expressed in the introduction about the
aims of a sentencing support system: The primary object of such a system should
not be to help a judge attain some superficial measure of uniformity, but rather
to consider uniformity in addition to the various sentencing policies explicitly
formulated by legislators or otherwise.
Ashworth (1992) considers three techniques for reducing sentencing disparity,
and is thus very relevant to our discussion. These techniques are: (i) Statutory
sentencing principles, (ii) Judicial self-regulation, and (iii) Numerical guidelines.
(i) The first of these approaches is not relevant to Israel, but will be considered
for the sake of completeness.
The idea behind this approach is to declare the leading principles of sentencing
in legislation, but leave the judges to apply and to individuate these principles.
Sweden and England are examples of countries in Europe, where this approach has
been implemented. The penal code of Sweden identify Desert4 as the primary basis
of sentencing (von Hirsch 1992). A similar approach in England was formulated in
(Home Office White Paper 1990)5 from which we shall bring two excerpts:
The first objective for all sentences is denunciation of and retribution for the
crime. (Home Office White Paper 1990, section 2.9).
Deterrence is a principle with much appeal ... But ... It is unrealistic to con-
struct sentencing arrangements on the assumption that most offenders weigh
up possibilities in advance and base their conduct on rational calculation.
(Home Office White Paper 1990, section 2.8).
(ii) The second approach (Judicial self-regulation) is directly relevant to our
discussion, as it forms the basis and justification for our case-based system. It will
be considered in Section 6.2.2.
(iii) The third approach (Numerical guidelines) was briefly mentioned in Sec-
tion 1.3. It is not relevant to Israel, but a computer system has been built imple-
menting this approach in USA. It will be studied in the next section.
3. Non-Intelligent Systems
3.1. ASSYST
An extreme approach towards attaining uniformity in sentencing has been taken in
the United States. The Sentencing Reform Act of 1984 is the statutory basis for the
present Federal Sentencing Guidelines, 6 which became effective in 1987. These
guidelines are very specific, defining ranges of sentences covering felonies and
some misdemeanours. If a sentence passed fails outside the prescribed range the
4 Desert is often used for Retribution. Roughly speaking Desert is Retribution without Vengeance.
5 Later adopted by Parliament in the Criminal Justice Act 1991.
6 There are significant differences among the federal guidelines and others adopted by many
individual states. However, those differences need not concern us.
[53]
158 URI J. SCHILD
offended party may appeal, solely on these grounds. The United States Sentencing
Commission, established in 1985, has the ongoing responsibility for the devel-
opment and monitoring of the Federal Sentencing Guidelines, and recommends
appropriate modifications to the United States Congress.
We shall not discuss the detailed reasons for the adoption of this approach,
except that it relates to a general disappointment with the rehabilitation approach
in force until then. When an American judge sentenced an offender to a prison
term, he almost never had any idea of how long that offender would actually be
behind bars (Reynolds. and Beck 1993). It might not even be possible at the time
of passing sentence to foresee how long that would be, as it depended upon the
future behaviour of the offender. Under such circumstances the actual prison term
of offenders who committed more or less the same crime could actually vary from
one year to twenty. For early critique of the rehabilitation approach and of the
sentence-guidelines approach now in force see (Martinson, 1974; Hudson, 1987).
Later papers criticizing the guidelines approach are (Ashworth 1991; Tonry 1993).
Sentencing guidelines reduce the number of factors (and their combinations)
taken into account. Obviously disparity in sentencing is reduced, for different courts
will treat like cases alike. However, it may also be argued that the courts are forced
to treat as alike cases that ought to be treated differently.
The approach to consistent sentencing as embodied in the sentencing guidelines
has met with strong criticism from the American judiciary. It is felt that very little
latitude is left for the judge in his decision, thus severely constraining any wish he
may have to consider the particular circumstances of the case at hand. Judges have
actually resigned rather than being forced to apply the restrictive guidelines (Freed
1992; Tonry 1993a).
A computer system called ASSYST (Simon and Gaes 1989) is based on these
federal sentencing guidelines. It elicits from the user all the information required
to make a decision, i.e. to determine a sentence according to the guidelines. It is a
rather complex algorithm described in a general way in (Simon & Gaes, 1989). The
authors of that paper mention that detractors have called the program a "Sentencing
Guideline Calculator" (Simon & Gaes 1989, p. 199). That, however, should not be
interpreted as a criticism of the program itself. ASSYST appears to be a profes-
sional, user-friendly and robust program for doing exactly what it is intended to
do: Assist judges (or rather probation officers, as it appears) in applying the federal
sentencing guidelines. The criticism is directed at the guidelines themselves, which
make the sentencing process rather mechanistic.
The use of ASSYST in the federal courts has been reported as very successful.
However, it is of no relevance in justice systems that do not subscribe to fixed
sentencing guidelines.
[54]
3.2. LIST
This system was developed at the University of British Columbia (Hogarth, 1988).
It takes a step in the direction of supplying a judge with relevant information,
without actually prescribing a sentence. The database consists of sentencing de-
cisions of the Provincial Court, Supreme Court and Court of Appeal of British
Columbia. The user interface of the system uses menus extensively and operates in
the following way:
(1) The user selects (from a menu) the appropriate legal domain (Criminal Code
of Canada, Narcotics Control Act, etc.).
(2) He next selects the particular offence from a list of offences.
(3) He now defines a small number of offender characteristics. For robbery these
would be: age-range, use of weapon (yes/no) and past record of violence
(yes/no).
After thus having specified the facts of the case at hand the user may proceed in
one of several ways.
(i) He may request a histogram providing types of sentence ranges on the vertical
axis for all cases in the database which match the facts of the case at hand. He
may also request a table of the individual dispositions.
(ii) The user may view all or some of the relevant decisions of the British Columbia
Court of Appeal. These cases are retrieved according to the type of the sen-
tence.
(iii) The system permits the user to retrieve cases according to aggravating and
mitigating factors relating to the offence, the record or the offender.
(iv) The system allows the retrieval of a summary of British Columbia Court of
Appeal sentencing decisions which lay down general principles or specific
factors which guide the exercise of discretion.
The statistical knowledge supplied by this system is said to be of great impor-
tance for a judge in order to determine what may be called an 'average' sentence
with respect to the facts known to the system.
However, this approach has also met with some criticism. The statistical knowl-
edge embodied in the Canadian system is based on a small number of characteris-
tics as mentioned above. This does not suffice to express the actual complexity of
the sentencing process. To quote a paper dealing with computer aids for sentencing:
Without being too imaginative, the following dimensions could be used to de-
fine 'criminal record': (a) the number of previous convictions, (b) the recency
of the last conviction, (c) whether the past record includes violent offences,
(d) the length of time since the offender first was convicted, (e) whether the
present offence was more serious than the most recent offence he had been
sentenced for, and (f) the nature or severity of the offender's most recent
sentence. It does not take a mathematical wizard to realise that if there are
[55]
160 URI J. SCHILD
even as few as three or four levels of each of these six variables, there are over
700 combinations of aspects of this one variable - criminal record. (Doob &
Park 1987, p. 611)
It is well-known that there are many more variables like criminal record, each
with several dimensions that in tum have different levels. The total number of
combinations of all parameters would reach into the tens of thousands. A judge
would first have to determine the values of all those parameters in the present case.
This calls for much discretionary effort, and may easily discourage the judge. Next,
statistics would have to be compiled for similar cases. But the number of previous
cases corresponding exactly in all parameters to the case at hand would be so small,
that the statistical data would not be significant.
Significant statistics could of course be compiled over a very long period. But
this would be of little use: Sentencing data from, say, twenty years ago would
probably not be relevant to the sentences being given today. Finally, detailed data
relating to sentencing parameters is simply not available. We are not aware of any
country where more than a few sentencing parameters are recorded and stored for
future use.
Lovegrove is a criminologist who has worked extensively in the area of sen-
tencing and sentencing models. He has argued that there are actually two problems
with Hogarth's approach:
First it is acknowledged that there is disparity in sentencing, and it is important

that a few disparate cases should not appear to be the norm. Second, because
any classification scheme could not cover all the relevant case factors and their
variations, the sentences for the cases falling within anyone particular cell
would cover the range. But where there were only a few cases representing a
cell, one would have little idea where the cases lay in that range. (Lovegrove
1989,42)
The question arises whether a judge passing sentence actually determines all
those thousands of parameters for each case. It is obvious that the human brain
works differently than a computer. It seems to be able to estimate the overall picture
without a detailed analysis of each single parameter. This is the essence of human
discretion, but a computer must take the detailed approach.
A different approach to a statistical sentencing system is proposed in (Doob
& Park, 1987) depending on a much larger number of variables. According to
(Lovegrove, 1989) this system nevertheless suffers from the same limitations as
Hogarth's system. Furthermore, it calls for SUbjective estimates of several factors
for determining the seriousness of the crime by the user, which is a "recipe for a
system designed to foster and hide inconsistency of approach" (Lovegrove 1989,
47). Finally, why should a judge be bothered to apply discretion in order to deter-
mine various factors to be input to an advisory system? He might as well apply
discretion directly to deciding a sentence.
[56]
The conclusion is, that the statistical data defining the starting point in the de-
liberations of a sentencing judge, while undoubtly superficially impressive, may
simply be misleading. It is therefore quite interesting and somewhat surprising that
other systems based on statistical information have since been built.
3.3. THE AUSTRALIAN (NEW SOUTH WALES) SYSTEM (SIS)

An Australian system: SIS (Sentencing Information System) has been built with
the aim: "to promote sentencing consistency through the dissemination of infor-
mation", and assuming that: "The provision of information on sentences imposed
by the courts was seen to be a more acceptable method of achieving consistency
than the establishment of a sentencing councilor the development of sentencing
guidelines". (Chan 1991, 139).
Values of the following variables must be input by the user:
1. Jurisdiction (local or higher court).
2. Act, section number or other specifications of the principal offence.
3. Whether other offences are taken into account in the sentence for principal
offence.
4. Number of counts of principal offence.
Other offender characteristics may optionally be supplied by the user:
5. Prior criminal record.
6. Whether the offender was on ball or bond at the time of the offence.
7. Plea entered.
8. Age of offender.
The prior criminal record is chosen among four categories: (a) No prior convic-
tions, (b) No prior convictions for an offence of the same type as the principal
offence, (c) At least one prior conviction for an offence of the same type, (d)
At least one prior conviction for an offence of the same type, which resulted in
imprisonment. Type refers to one of the following categories: (1) Drug offences, (2)
Sexual offences, (3) Fraud-dishonesty, (4) Property offences, (5) Offences against
the person, (6) Driving offences, (7) Breach of order/escape.
The system supplies statistical data as well as other information similar to the
output obtainable from the LIST system.
The builders of the system (Weatherbum et aI., 1988) are well aware of Love-
grove's criticism of Hogarth's system quoted in the previous section, which they
have attempted to answer:
Since the statistics involve the universe of cases rather than a sample, the
question of a small number of cases is immaterial. The absence or paucity
of cases in a particular cell simply reflects the level of available precedents.
(Chan 1991, 148 citing Weatherbum)
[57]
162 URI 1. SCHILD
We have great diffkulty accepting this argument. It is in any case not relevant
whether a statistician would agree to it. The fact is, that judges are not experts
on statistics and may misinterpret the data and give it far more weight than is
scientifically correct. Many additional points of criticism of the system are found
in (Greenleaf, 1991).
3.4. THE SCOTTISH SYSTEM (SIS)

A Scottish system also named SIS (Sentencing Information System) has been de-
scribed in (Hutton et a!., 1995). This paper makes no mention of the Canadian and
Australian systems, nor does it relate to the criticism that has been voiced of the
statistical approach. Its data-base contains all Scottish High Court convictions over
a five year period (4500). The case-base thus relates mainly to first instance sen-
tences, but the user may request information about possible appeal and its outcome.
The case-base may also be restricted to appeals cases.
This is not a case-based system in the sense that it performs case-based reason-
ing (CBR). It retrieves cases from the case-base according to certain indexes and
compiles sentencing statistics, using a relational data-base environment (SQL). For
an overview of relational databases, SQL and their use in the legal domain, see
(Zeleznikow & Hunter 1994,29-38).
Initially the indexing addressed the important features in each offence cate-
gory. However, in order to overcome the problem of insufficient data for statistical
purposes convictions were reclassified using composite similarities (Hutton et aI.,
1995, 91). The latter were identified by discussion with the judiciary. Thus the
problems formulated by Lovegrove and described in Section 3.2 also appear here.
The system operates in the following manner:
1. The user selects offence category and the classification of the offence within
that category.
2. He may optionally choose offender characteristics and/or offence characteris-
tics.
The distribution of sentences are then presented (as a histogram) on the basis
of disposal: Imprisonment, probation, fine, etc. Each of 15 categories may then
be further explored to show a distribution within that category, yielding details of
sentence, reference to actual cases with low level details.
The prototype was used by four high court judges. They were especially inter-
ested in the potential of the system for informing sentencing decisions on unusual
cases (Hutton et aI. 1995, 93). It also appeared that the judges were particularly
interested in displaying low-level details of cases, in order to examine their rele-
vance to the case at hand. It is thus possible that the judges were alert to the danger
of having a collection of 'extreme cases' in a small sample. A production system
has now been completed, and is scheduled to be introduced in the Scottish courts
within a short time (Wilson, 1996).
[58]
3.5. THE ISRAELI SYSTEM?

Despite the fact that the sentencing situation in Israel is fundamentally different
from the situation in the USA (described above in Section 3.1), there are neverthe-
less people (mostly politicians) who believe that sentencing guidelines should also
be introduced in the Israeli courts.
The previous president of the Supreme Court of Israel decided before his recent
retirement, to initiate the development of a statistical information system by the
Center for Computers and Law at Bar-Han University. This created a dilemma for
us. On one hand the arguments given in the previous sections show that the use of
such a system is scientifically suspect (to say the least), and our academic integrity
should not allow us to participate in such a project. On the other hand, we did not
want to jeopardise our good relationship with the Israeli judiciary, which has been
very helpful in our CBR project described below.
After taking advice from criminologists, expert lawyers and judges, it was de-
cided that an appropriate statistical sentencing information system could partially
overcome the various defects described previously. This could be done, not neces-
sarily by adding large amounts of detailed information about the case at hand, but
by examining the case-base along one additional dimension only, the dimension of
criminological sentencing approach.
The four classical principles play an important role in the study of sentencing.
Thus (von Hirsch and Ashworth 1992), which is a respected collection of papers
that consider the issue of uniformity and sentencing principles, is organized ac-
cording to these principles. It is even more important that these principles also play
a central role in the sentencing process itself. In chapter 2 we saw that legislation
in some countries explicitly specify which of the four sentencing aims are to be
considered central and which are less important. Furthermore, according to the
saying by Lawton L.J. quoted in Section 2.1 every judge should consider these
sentencing aims when passing sentence.
It is thus understandable that developers of the Israeli systems described below
decided to incorporate these sentencing aims as basic parameters in their sys-
tems. However, there is absolutely no empirical data (or theoretical basis) for
assuming that different sentencing approaches is the only or even the major reason
for sentencing disparity. Ashworth (1995, pp. 60-80) discusses many other aims
and principles that conceivably could be as important for a statistical analysis as
the parameter based on the classical sentencing aims. In other words, adding this
parameter is not necessarily an answer (even a partial one) to Lovegrove's criticism.
As in other systems the user must initially supply the following basic informa-
tion about the case-base:
1. The relevant court(s)
2. The relevant paragraph( s) of the criminal code
3. The time period of retrieved cases (one or five years).
7 It has not yet been named. By induction it should of course also be named SIS.
[59]
164 URI J. SCHILD
Information based on cases retrieved according to these parameters only is of

extremely dubious value, as we saw above. Yet, the judiciary insists upon receiving
it, and an existing prototype of the system will supply averages and histograms
based only on the above information.
For a statistical sentencing base and given values of the above three parameters,
the concept of an average can be defined in different ways:
1. The mean sentence is the sum of all retrieved sentences divided by their num-
ber (relates only to custodial sentences or other sentences with a numerical
value).
2. The median sentence is the sentence dividing the set of retrieved sentences into
two sets of equal size.
3. The mode is the sentence (or rather sentence-interval) appearing with the great-
est frequency among the retrieved sentences.
Various kinds of statistical analyses may in principle be applied to the retrieved
data. However, it has heen argued by the judiciary that judges would not appreciate
the results nor would they have time to examine them.
The user may request more precise information by checking one of the follow-
ing items, hopefully leading to a reduction of the case-base which will yield more
meaningful statistics:
Question: Sentencing approach was:
Rehabilitation
Deterrence of offender
Deterrence of the public
Isolation of offender
Retribution
Again, appropriate averages and histograms will be presented to the user.
It was decided not to include case-descriptions at all. The relevant information

needed to build a statistical database for the system may therefore be extracted
from the central police files of offenders, which are continuously being updated.
The actual computer program is therefore extremely simple (compared for example
to the use of relational data-base techniques employed by the Scottish system), as
no single case has to be retrieved.
Not enabling the user to access any descriptions of cases from the case-base
appears to be a most serious defect. In the other systems the user can always check
whether the facts of a particular case in the case-base are similar to the facts of the
present case. Thus the Scottish judges who tested the prototype, actually looked at
the individual cases. However, as one of our judges has pointed out, there may be
a difference between a user's behaviour when experimenting with a system under
test-conditions and the way he will use the system under realistic work conditions.
[60]
Judges may not find the time to read old records when faced with a stack of new
cases to decide.
The police files do not at present contain any information on the sentencing
approach. It was therefore decided that in future cases, judges will be asked to
supply their answer to the above question, which will become part of the official
protocol of the case. The scientifically dependable part of the program (if one can
call it that) can therefore not become operational before sufficient statistical data
has been collected.
It is intended that the database for the system will be built up from the present.
No judge can of course be forced to answer the above question, but only those cases
which do have answers will be selected and incorporated in the statistical database.
Thus judges that choose not to answer these questions will know that they have no
influence on future sentencing policy, and that their decisions will not be reflected
in the statistical data available in future cases. It is hoped this will be sufficient
incentive for the judges (or their clerks?) to supply answers to the relevant question
in the appropriate part of the case-protocol.
Deplorably, it is thus known a priori that the system will have some serious
defects:
(1) The lack of capability to check the actual facts of cases contributing to the
statistics. (Judges may not find the time to use such an option, but it ought to
be present).
(2) The lack of statistical confidence in the general results, retrieved without the
sentencing approach parameter.
(3) The uncertainty whether the data relating to the sentencing approach parameter
will ever be compiled.
(4) Despite these three defects one could perhaps hope that the use of the system
by the judges would limit the disparity of their future sentencing. However,
the present database reflects the current non-unifonn sentencing situation, and
its data is not meaningful. This is essentially Lovegrove's first argument in
Section 3.2.
A prototype based on fictitious data has been built, and funds have been al-
located for a real system. We hope that time and costs of development of the
real system will be within the projected bounds. Collection of data relating to the
sentencing approach parameter will depend on the cooperation of the judiciary.
3.6. CRITIQUE OF STATISTICAL INFORMATION SYSTEMS IN THE SENTENCING

DOMAIN
Two such systems were developed (and abandoned) in Canada. Two other systems
are actually functioning, and a third is under development. It is therefore important
to summarize in a precise manner what the problems are with this kind of system.
[61 ]
166 URI J. SCHILD
1. Assume that the statistical data is retrieved based on only a small number of
parameters. Then data for violent "professional" bank robbers would perhaps
be considered together with the data for purse snatchers, giving a misleading
picture of the fine··structure of the sentences.
2. Assume on the other hand that a very large number of parameters are used
to retrieve the statistical data. The judge may not bother to input all those
parameters in the first place. Even if he (or his clerk) does, the number of
cases corresponding to given values of these parameters could be so small, that
the retrieved data would not be statistically significant.
3. It is difficult to understand how uniformity can be obtained based on non-
uniform data. A uniform database can be obtained by building it up from
scratch using uniform data only. It is not easy to see how this can be done.
4. Judges are not professional statisticians and may attach too much weight to
data than is scientifically correct. Users of a statistical computer system who
are not expert statisticians or computer scientists, often trust facts and figures
supplied by a computer, and automatically accept them as significant.
The principal source of guidance on sentencing (apart from legislation) consists
of previous judicial decisions. Statistical systems seem to be a convenient way of
summarising such material. As we have seen all such systems have serious defects,
that cannot be overcome. They may actually supply misleading results, as users
who are not expert statisticians or computer scientists, often trust facts and figures
supplied by a computer, and automatically accept them as significant. However,
such systems give decision support in a compact and readily surveyed form.
Additional material on statistical information systems is found in Section 7.
4. Intelligent Systems
4.l. RULE-BASED EXPERT SYSTEMS
Historically the first knowledge-based systems were rule-based, and it seems rather
surprising that this technology was not applied to the problem of criminal sentenc-
ing long ago. In the general area of Al and Law many rule-based systems have of
course been constructed. An overview of such systems is found in (Sergot, 1991),
while (Susskind, 1987) gives a most thorough jurisprudential analysis of rule-based
expert systems in law.
Apart from ASSYST, which could be rewritten as a rule-based system (but
without heuristics), two systems are relevant, and will now be described.
4.l.l. The System for Probation Officers

A system has been built in Israel for the use of probation officers in recommending
sentences for young c.iminals (Shapira, 1990). This system has been in existence
and actual operation by the Youth Probation Services for several years. While
technically not a rule-based system in the sense that it has a separate knowledge-
[62]
component in rule-form, it should nevertheless be classified as such. For the expert

knowledge was originally formulated by the human experts in something close to
rule-form. The sharp distinction between the knowledge-base and the engine ofthe
system was lost during the translation to a programming language (DBase III).
The probation officer does not necessarily have to adopt the recommendation
of the system, but she must give a detailed justification for any deviation from it.
Of course, even if the probation officer adopts the recommendation of the system,
the judge does not have to follow this recommendation, though she usually does
(in 82% of the cases: (Shapira 1990, 161). The system has met with great success,
it has been readily accepted by the probation officers, and is also said to save time
and improve their efficiency.
Here we have an example of a knowledge-based system that actually assists
in the sentencing process. It is, however important to consider the environment
where this system operates. Probation officers are governmental employees who
do not enjoy (and do not expect) the same kind of independence as judges. They
must carry out orders and policy as formulated by their superiors with or without
a computer system. If this policy happens to be expressed in rule-based form and
included in a computer system, it makes their work easier, as experience indeed has
shown. Furthermore, these probation officers are employed by the youth courts.
Sentencing policy in those courts is much simpler than in courts for adults, and no
conclusions about a similar system for adults can be drawn.
We mentioned above that the explicit form of the rule-base was lost in the actual
implementation. It is usually considered a great advantage of rule-based expert
systems, that the user may examine the applied rules, and thus convince himself if
necessary that the answers supplied by the system are appropriate. This process is
thus not possible in the probation system. We have interviewed several probation
officers, who actually use the system, and asked whether they were bothered by
this. From their answers it appears that, based on long experience, they implicitly
trust the system and the answers it supplies.
4.1.2. Bail Advisor

Work is in progress in England on a rule-based system for bail-setting decisions
(Has set, 1993). It appears that decisions on whether to release a suspect on bail
lack uniformity. While one magistrate (judge) will deny bail to an arrested person,
another judge will set some minimal amount of bail for a suspect detained under
almost the same circumstances.
On one hand the financial expense of remanding prisoners is very high, on
the other hand there is of course great concern about persons who commit of-
fences while free on bail. The criminological aspects of the project have also been
described in (Burrows, Henderson & Morgan, 1994).
The English system under development uses the following approach. The mag-
istrates are interviewed about the way they go about deciding about release on bail.
This information is expressed in rule-form with the intention to build a rule-based
[63 ]
168 URI J. SCHILD
expert system. When a magis-trate has to make a bail-release decision he would

query the system, which would supply him with a definite answer to the question
whether to release or not to release the suspect. Obviously the magistrate would not
be forced to follow the decision of the system, but hopefully this decision would
be an important guide for him, when handing down his decision.
The applicability of this extremely interesting system is at present inconclusive
(Hasset, 1994). Also here it is important to consider the environment where the
system is supposed to operate. Decisions on bail in the U.K. are made by the lowest
level of judiciary, the magistrates. These are laymen of impeccable background,
who have volunteered for this job. Thus, on one hand this group exhibits certain
similarities with the probation officers, on the other hand the magistrates have many
of the characteristics of professional judges.
4.1.3. Other Work

Berman & Hafner (1989) surveys the problems related to computer systems for
sentencing. The paper also discusses ways of actually implementing rule-based
expert systems for sentencing, and raises the possibility of using MYCIN-like
weights. MYCIN (Davis et aI., 1977) is one of the first medical expert systems
developed. It associates a so-called certainty factor (Johnson & Keravnou, 1985)
with each rule and combines these factors according to the laws of fuzzy logic.
Berman & Hafner (1989) proposes to associate 'danger factors' to rules in such a
way that a ')ail term may be indicated when the danger factor exceeds a particular
value" (p. 935).
The weighting approach of MYCIN and actually of most classical expert sys-
tems has not met with uniform approval (see Wellbank (1983)). Furthermore, ex-
perts in the legal domain do not usually reason in terms of numerical weighting
schemes or probabilities (Ashley and Rissland 1988, p. 239), even though the final
outcome of such reasoning in the case of sentencing can be numerical (e.g. length
of prison sentence).
The JURICAS shell (van Noortwijk et aI., 1991) has been used to create an
advisory system in the area of remanding prisoners in custody.
4.1.4. Critique of Rule-Based Systems in the Sentencing Domain

The practical usability of rule-based computer systems in various domains has been
under discussion for over twenty years, with opinions ranging from total rejection
to total acceptance. We have surveyed two such systems above. The first of these
has had a proven success, the second is under development.
Both these two systems deal with decision problems similar to criminal sen-
tencing. However, the persons making those decisions are not professional judges.
In the Israeli system they are social workers (youth probation officers), and must
necessarily follow the regulations of their service. In the English case the lowest
level of judiciary is made up of non-professional judges. They do not have to
[64 ]
follow any strict regulations imposed from above, and thus have the independence
of judges. It is possible that they will agree to work with a rule-based computer
system.
This, however, is not the case with professional judges. Our impression from
speaking with judges is one of total rejection of the idea of using fixed rules for
sentencing. The rules of a sentencing system would reflect the private view and
opinion solely of the judge formulating the rules. These opinions may not neces-
sarily be held by other judges. A panel of judges may not reach an agreement about
sentencing rules, and even if they did, no judge would be obliged to follow those
rules, or even to consider them while passing sentence. Adopting a set of rules by
legislation is of course equivalent to the American approach through sentencing
guidelines.
We may add some further reasons why the rule-based approach is not appro-
priate for sentencing advisory systems. These reasons relate to rule-based expert
systems in general, and have been observed many times in the past (Wellbank,
1983).
(i) Experts are often inconsistent in the sense that they do not practice what they
say they do. In other words, even if they are willing to formulate rules, they do
not always follow those rules themselves.
(ii) Rules extracted from a panel of experts are often conflicting. Even a single
expert often contradicts himself. In the case of judges these inconsistencies
may reflect the actual inconsistent sentencing.
(iii) A prominent feature of rule-based systems is their capability of explaining the
rests they supply when queried. However, the systems do not enable differ-
entiation between rules of a technical nature and rules with conceptual expert
knowledge, so the explanation of results becomes just an explanation of the
formal deductions. Motivation for the application of specific rules is often
lacking, as the systems do not have deep domain knowledge. (See, e.g. (Alty &
Coombs 1984, 101-102; Davis & Lenat, 1982) on the expert system TEIRE-
SIAS, or (Chandrasekaran & Mittal, 1982; Clancey, 1983)). We have seen for
the probation officer system, that the capability of giving explanations is not
critical. This cannot be true for judges. Should one really build a rule-based
sentencing system, explanation of the conclusions must be perhaps the most
prominent feature.
(iv) Extracted rules have semantic vagueness. It is not clear how they should be
formalised in a computer program (Schild 1992, pp. 133-137). Whatever for-
malisation is decided upon, implies that discretion is ignored.
(v) The explanation construction problem is computationally intractable. Building
a completely new explanation from scratch is a very time-consuming process.
Explanations are essentially causal chains. Small chains are reasonably easy
to build, but the number of possible chains gets very large as the length of
the chain goes up. For explanations the size that humans routinely create, the
search space that a system that built such explanations from scratch would have
[65 ]
170 URI J. SCHILD
to serve is very large. If we had to build all (or even most) of our explanations
from scratch, we would never have any time to do anything else (Schank,
1994). The case-based and model-based reasoning paradigms were developed
as methods for bypassing reasoning from scratch. They will be considered in
the next two sections.
4.2. CASE-BASED SYSTEMS

Case-Based Reasoning (CBR) is a problem solving approach by humans and com-
puters where: "new problems are approached by remembering old similar ones and
moving forward from there" (Kolodner 1993, xiv). Given a new problem, a CBR
program (i) retrieves relevant cases from a case-base (using appropriately defined
indices), (ii) chooses the most similar case, and (iii) adapts its solution to the new
problem. The first step of CBR is often called Conceptual Retrieval.
There are some domains and tasks that are so well understood and well-defined
that a small number of rules suffice to reason about them. Exhibiting the deduction
process to the user would convince him about its validity, and some weighting
scheme could also be applied. CBR would probably not be appropriate for such
areas. It should be clear by now, that the sentencing domain does not satisfy these
criteria, and CBR is therefore a good candidate for the underlying paradigm of a
sentencing support system.
Several researchers. have previously applied methods of CBR to the legal do-
main, but not to the area of sentencing (see, e.g., Rissland & Ashley (1987); Ashley
(1990); Branting (1991); Skalak & Rissland (1992». Ashley (1992) is a most
comprehensive overview of the use of CBR in Law.
An early CBR program relating to sentencing was JUDGE (Bain, 1986). It used
the sentencing domain to test cognitive theories of reminding and problem-solving.
It was not intended as a program of practical use by judges. Our experience from
interviewing judges closely corresponds to the cognitive observations in Bain's
work (best described in Riesbeck and Schank (1989»:
These observations may be summarized as follows (Riesbeck and Schank 1989,
pp.98-110):
Prior record:
1. Judges try to predict the likelihood that an offender will repeat.
2. A judge will treat an offender more harshly if he believes that there is reason
to predict recidivism.
3. Judges base their predictions about an offender's future behavior in part on the
presence and severity of a prior record.
Forming explanations:
4. Judges seek reasonable explanations for criminal behavior.
5. Judges base their expectations about the arguments of the attorneys in a case
on their own explanations of the offender's behavior.
[66]
The Victims:
6. A judge tries to punish an offender to an extent which is consistent in part with
the amount of harm caused to the victims.
Experience-based generalizations:
7. The extent to which a judge considers certain features of cases and of offenders
to be significant is a function of recent experiences he has in dealing with those
features.
8. The extent to which a judge considers certain features of cases and of offenders
to be significant is a function of prototypical experiences he has in dealing with
those features.
Choosing a Structure for Sentencing:
9. Judges enumerate the set of choices which they believe they could impose as
sentences.
10. The more a judge believes that an offender is likely to repeat a crime which
involves victims, the more he will prefer to remove the offender from society
by giving him a prison term.
Conditions on Choices:
11. Judges attach conditions to the sentencing choices which they enumerate.
Motivation of the Offender:
12. The process of explaining an offender's behavior involves inferring his moti-
vations.
Case-based Reasoning:
13. Judges sometimes formulate sentences by using case-based reasoning.
14. When a judge applies a sentencing strategy from one case to another, he'll
modify the sentence to accommodate feature differences between the cases.
Rationalization:
15. Judges stress the relative importance of specific features of cases to support the
rationality of their sentences.
Indexing and Retrieval:
16. Judges retrieve cases from memory - their own cases as well as others - to
compare to new cases.
17. Judges relate causal differences in cases to differences in the sentences which
they give.
Some further conclusions about the behavior of judges when passing sentence
are found in (Riesbeck and Schank 1989, pp. 111-113).
Further discussion of CBR-systems for sentencing, including the description of
a system we have developed, is given in Sections 5 and 6, where we shall return to
Bain's work.
[67 ]
172 URI 1. SCHILD
4.3. MODEL-BASED SYSTEMS
The rule-based systems dealt with in Section 4.1 have a knowledge-base in rule-
form, with rules that have been elicited from human experts. It is also possi-
ble to create a computer system, where the knowledge is based on a theoretical
model. The theoretical knowledge may be expressed in rule-form or in any other
convenient manner. It may be combined with other kinds of knowledge (e.g., a
case-base), or appear in a stand-alone system. A general overview of model-based
systems is Forbus (1988).
In the area of knowledge-based systems there are examples of the success of
model-based ones, especially in combination with other kinds of knowledge-bases,
e.g., case-bases (Kolodner, 1993). Criminologists actually build sentencing models
(Wilkins et al. (1978): Fox (1981); and Lovegrove (1989)), though they may not
agree among themselves about the validity of those models. So it should in princi-
ple be possible to build model-based computer systems for sentencing. DeMulder
and Gubby (1983) describes a working system in the sentencing domain using a
model-based approach (the Hulsman model), and an extremely interesting eco-
nomic model of sentencing is described in Posner (1985). However, the judges do
not believe in these models or trust them - as the criminologists are the first to
acknowledge themselves. We shall therefore not discuss this kind of system any
further.
We emphasise that our rejection of model-based systems in no way reflects on
the accuracy of those models, which we have no intention - indeed no qualification
- for analysing. In the introduction we stated that our object is to examine systems
that judges would actually use. Model-based systems do not appear to fulfil that
criterion.
4.4. OTHER PREVIOl"S WORK (INTELLIGENT SYSTEMS)

A small number of other projects have dealt with computer systems and sentenc-
ing. We shall not consider these any further, as the systems discussed above are
representative for our purposes. A system has been developed in Tennessee for
sentence calculation (Reynolds & Beck, 1993). It assumes that the judge has passed
sentence, and calculates the release dates for offenders. It illustrates how difficult
it is indeed to compute the actual term of imprisonment in the US, as mentioned in
Section 3.1 above. It is not of relevance to our problem.
5. Conceptual Case Retrieval

In the previous sections we have described the problems relating to sentencing
advisory systems, and previous work relating to such systems, including a brief
mentioning of case-based systems. In this and the following section we shall extend
the discussion of the latter systems, including our own contributions.
[68]
5.1. BACKGROUND ON RETRIEVAL SYSTEMS
The first step of the case-based paradigm as considered in Section 4.2 is case re-
trieval. Textual information retrieval systems have been in existence and use in the
legal domain for over twenty years. A user will state his requirement by specifying
a word or a set of words connected by Boolean operators, and the system will
retrieve those textual documents which satisfy the requirements. Two concepts are
relevant to evaluating the retrieved results of a query (Salton, 1989)):
(1) Precision - the ratio of the number of retrieved relevant documents and the
total number of retrieved documents.
(2) Recall - the ratio of the number of retrieved relevant documents and the total
number of relevant documents.
Thus for a given query the precision and the recall should both ideally be one.
That would signify that all relevant documents and only those have been retrieved.
However, experience shows that the classical method of information retrieval can
be problematic. A user is often interested in locating documents relevant to a
certain concept. But texts and concepts may not have a one-to-one correspon-
dence due to homography and synonymy, and the search operators for specifying
relationships between words of text are crude approximations for specifying the
relationships between the underlying concepts (Bing, 1987).
Instead researchers have developed conceptual retrieval systems where the search
method is based on the meaning and significance of the retrieved documents: Con-
ceptual retrieval is the process by which systems index and retrieve information
using conceptual structures rather than text structures.
In the legal domain the term 'conceptual retrieval' has been used by several re-
searchers, primarily by Hafner (see Hafner (1981) and Hafner (1987)). According
to Hafner (1987) the knowledge-base of a conceptual retrieval system consists of
three parts:
(1) A domain knowledge model, which defines the concepts the system should
understand and know about, both legal and common-sense knowledge. Also
relations among these items of knowledge are included in the domain knowl-
edge.
(2) Individual case descriptors, which are descriptions of each case in the case-
database structured according to the concepts of the domain knowledge model.
(3) A hierarchical rule system (called the issue/case discrimination tree in Hafner
(1987)). This set of rules will enable the system to locate and retrieve the cases
relevant to the problem at hand.
Given a conceptual retrieval system with three components appropriate to the
sentencing domain and a sentencing case-base, a judge about to determine a sen-
tence could retrieve precisely those cases relevant to a stated concept.
However, the creation of such a system is far from easy. Each of the three
components specified above raises a different problem. With respect to the first
[69]
174 URI J. SCHILD
component: It is not possible to include all necessary legal knowledge and certainty
not all common-sense knowledge in any present state-of-the-art system. As for the
second component, the following view has been expressed by a leading authority:
If intellectual indexing terms are assigned to documents as part of a related
publishing scheme - as for WESTLAW and the Butterworths material in
LEXIS - one would obviously include them. But if the costs are not met by
such additional activities, one would hardly be able to justify the indexing
costs by increased performance. (Bing 1989, 154)
Finally, concerning both the first and the third component, expert knowledge
is necessary in order to create the appropriate hierarchical structures, but may not
always be available.
Besides Hafner's work an attempt at formalising cases for the purpose of con-
ceptual retrieval was carried out by Dick (1991). She applied Sowa's conceptual
graphs in the area of contracts, and each case was analysed using the Toulmin
argument model (Toulmin, 1958). Conceptual graphs have a well-defined seman-
tics, but the work necessary to formalise any given case is considerable. Dick only
carried out the analysis for a few cases. Mital et al. (1991) deals with conceptual
information retrieval in litigation support systems. Turtle (1995) is a comprehen-
sive overview of text retrieval, and Smith et al. (1995) is a recent paper describing
a system involving conceptual retrieval and automatic indexing.
Statistical support systems could in principle be classified as conceptual re-
trieval systems, as the first step of their operation is a selection of cases according
to some specified indices. They are knowledge-based systems in the sense that
the indices used for retrieval are determined through knowledge of the sentencing
domain.
5.2. OUR CONCEPTCAL RETRIEVAL SYSTEM

Keeping in mind the problems mentioned in the previous section, we have devel-
oped a computer system which uses conceptual retrieval in order to present relevant
information from a sentencing case-base. There is no overlap or connection be-
tween this system and our statistical information system described in Section 3.5.
The domain know ledge was elicited from an expert judge, the vice-President of the
Tel-Aviv District and Appeals Court, who has many years experience on the Bench,
especially in criminal cases. After many sessions and iterations the parameters that
judges take into account when passing sentence, were determined. These factors
were also found appropriate by our academic associates, a professor of law doing
research in the area of criminal law and a panel of criminologists.
Eliciting information from experts is always a difficult enterprise. Our method
was proposed by the judge himself. We decided, as it were, to create a questionnaire
to be filled out by criminologists building a sentencing case-base based on actual
trial protocols. In other words, the judge told us what a criminologist should look
[70]
for in the records, what issues and concepts relating to sentencing should be ex-
tracted and noted down. Obviously these concepts were the ones the judge himself
considered relevant to the sentencing deliberations.
The sentencing parameters were arranged as nodes in a set of discrimination
trees. One tree deals with features relating to the offender himself. In that tree, e.g.,
'not-main-offender' is a father-node of 'weak personality', which in its tum is the
father node of 'easily-influenced by others'. Other trees deal with the victim, the
offence itself, mitigating and aggravating circumstances, and general issues. Other
parameters relate to specific crimes.
It is obvious that the area of sentencing is associated with an enormous amount
of both common-sense knowledge and domain knowledge. In fact, a complete
model like the one proposed by Hafner (1987] would be almost impossible to
create, even if we limit ourselves to only a few specific paragraphs of the criminal
code. On the other hand, a hierarchical structure is both necessary and sufficient
for a system to carry out an initial step of retrieval. We therefore decided to use
the elicited domain knowledge only, without any additional common-sense knowl-
edge. The nodes of the discrimination trees were taken as indices to cases in the
case-base, i.e., these were the concepts a judge would be interested in.
We have limited ourselves to two serious crimes: Robbery and Rape. As in-
dexing of cases is a technically big problem for large case-bases, we decided to
establish a case-base spanning only the last five years. A larger time-span would
raise problem of trends and changing attitudes in sentencing, which at this stage we
do not intend to deal with. In Israel robbery and rape cases are heard in the District
Courts, and may be appealed to the Supreme Court. The maximum sentence for
each of these offences is twenty years imprisonment. This is one of the reasons
for choosing these crimes, as sentencing disparity becomes especially pronounced
when long custodial sentences are available.
For reasons to become apparent in the next section, we selected only Supreme
Court cases, and only those where the appeal relates to the actual sentence. The
prototype case-base covering the last five years contains less than a hundred cases,
and manual indexing was thus possible. It was carried out by a qualified criminol-
ogist and a graduate computer science student, who had also been involved in the
knowledge elicitation process.
The system assists the user to walk through the discrimination trees and check
the nodes that are relevant to his case. The system then retrieves those cases from
the case-base, which are indexed by the chosen nodes. Furthermore, retrieval is
also carried out for nearest neighbours of the chosen nodes. Thus, if the user has
checked the node 'not main offender' in the offender-tree, the system may also
retrieve cases indexed by a sister-node, e.g., 'assisting after execution of crime'.
Obviously cases retrieved by neighbouring nodes have less bearing than cases
which exactly fit the checked parameters. The system uses a Tversky-type simi-
larity metric based on the contrast model (Tversky, 1977) to order the retrieved
cases before presenting them to the user. For each case the system also presents
[71]
176 URI J. SCHILD
a list of the relevant parameters. The user can choose three levels of output: (1) a
formulation of the ratio of the case, (2) the parts of the law-report relevant to the
sentencing, (3) the entire law-report. Law-reports in Hebrew are not different from
law-reports in English. and we shall not show any output from the system.
The expert knowledge of the system is contained in its discrimination trees, in
its capability of retrieving not only directly relevant cases, but also cases indexed
by neighbouring nodes, which may have some bearing upon the new case. The
system was presented to several judges, who found it of far greater use than the
classical text retrieval systems, but of limited application because of the case-base,
which was too small and only contained appeal-cases. We observed that the judges
recognized most of the retrieved cases when testing the system. This is not surpris-
ing, as the case-base consists of Supreme Court decisions, most of which would
be landmark cases. We have already remarked in Section 3.5 that the readiness
of the judges to read the retrieved law-reports or excerpts from these under test-
conditions does not allow us to draw any conclusions about the applicability of the
system under realistic work-conditions.
The system could have been further developed in several directions. Using
Hafner's ideas much more intelligence could be added, and the case-base could
be significantly enlarged. That, however, would necessitate the addition of vast
quantities of domain-knowledge and especially of common-sense-knowledge. The
present state-of-the-art of Al does not allow that. So we did not proceed with any
further development of this system, and decided to use it not as a stand-alone
system, but as a complement to the system described in the next section.
6. Case-Based Sentencing Systems

6.1. INTRODUCTION
In previous sections we have described various approaches to sentencing support

systems. We shall now discuss the case-based approach in more detail and - as
an example - describe a system of our own. 8 Despite some theoretical foundation
for the CBR paradigm, systems that have been built over the years tend to be one-
of-a-kind systems, with only a minimum of common features. It is therefore not
surprising, that there should be several possibilities for implementing CBR-systems
for sentencing support. Ours is definitely not the only possible method, and we shall
therefore not dwell on technical details of the system in this paper. What we find of
importance is to argue that the CBR paradigm seems to be the only really feasible
one which may easily be accepted and actually used by the judges. Our system is
an illustration only of how this may be done.
8 This is the third Israeli system mentioned in this paper. The first one is the statistical information
system described in Section 3.5 and the second one is the conceptual retrieval system described in
Section 5.2. The second s)stem was eventually incorporated in the third one.
[72 ]
We remind the reader that our desire is to build a computer system that may
support a judge in deciding which sentence to hand down in a new case. Such a
system should not supply its user with a single, definite answer (i.e., a proposed
sentence). No judge would appreciate that. We propose an intelligent decision
support system, where several approaches and ways of reasoning will be produced
for the user, but the final choice will be his only.
6.2. JURISPRUDENTIAL FOUNDATION

Two interesting concepts from the domain of criminal sentencing are relevant to the
building of a computerised sentencing support system. One of these is the concept
of the 'tariff' as it appears in the writings and analyses of D.A. Thomas. The other
relates to the English guideline sentences (as opposed to the American sentencing
guidelines) handed down by the Court of Appeal.
6.2.1. The Tariff

It is traditional practice by the courts to establish a sort of scale - the tariff - which
grades the severity of the penalty in accordance with what is regarded as as the
gravity of the offence, without any individual consideration of the offender. Tariff
sentences (if actually imposed) are fixed at points on this scale relative to the worst
imaginable case of the offence in question, for which the maximum sentence is
reserved.
The concept of a tariff is never given an exact numerical definition, though it is
often used by writers on sentencing policy. Its meaning is best understood from the
following quotes from the person who is considered one of the foremost authorities
on this subject (Thomas, 1979; Thomas, 1982):
A tariff sentence will be imposed when the sentencer wishes to emphasise to
the public the gravity of the offence, while an individualised measure will be
chosen where the object is to influence the future behaviour of the offender.
(Thomas 1979, 14)
The principles of the tariff constitute a framework by reference to which the

sentencer can determine what factors in a particular case are relevant to his de-
cision and what weight should be attached to each of them. Properly used they
offer a basis for maintaining consistency in the sentencing of different offend-
ers, while observing relevant distinctions, making appropriate allowances for
individual factors and preserving adequate scope for the exercise of judicial
discretion. (Thomas 1979,29)
Thomas considers a number of offence categories. Within each category he
assumes a variety of factual situations may occur. For each such situation he sets
out the effects on the sentence. As we have concentrated on the offences of robbery
and rape, we shall bring material relevant to these crimes exclusively. The section
[73 ]
178 URI J. SCHILD
on rape (Thomas 1979, 112-117) includes the following (all references to the large
number of actual cases given by Thomas have been deleted):
The scale of sentences extends from a usual minimum of two years to an

upper limit of twelve years, although sentences both shorter' and longer are
occasionally seen. The factors which advance the sentence within this scale
include the degree of violence or threatened in the course of committing the
act, the infliction of other forms of sexual abuse, the involvement of more
than one defendant and the forcible abduction of the victim or the invasion of
the victim's house. Conduct on the part of the victim which increases the risk
of rape, such as willing participation in minor sexual activity or (to a lesser
extent) imprudent behaviour such as accepting a lift in a car from strangers,
is usually treated as a reason for some reduction in sentence. The age of the
victim does not appear to have any substantial significance (except that the
rape of a young child or an elderly woman may be an indication of the need
for psychiatric investigation). The victim's previous sexual experience is not
a major factor. The Court has said that 'it is not to be thought that whenever
the victim of a rape is a woman of experience or of dubious moral standards
it is any less an outrage to take that which the woman is not willing to give',
although it is relevant to consider the extent of any physical or traumatic injury
caused by the offence. The fact that the victim has forgiven the offender is not
necessarily relevant. (Thomas 1979, 113)
One could possibly extract a set of rules for sentencing using Thomas' material,
though quantitative information must be added. The general problems mentioned
above in connection with rule-based systems would of course also apply here. In
addition, rules based on Thomas' analysis could only be considered as secondary
legal material.
But one can also imagine another kind of computer support system based on
Thomas' analysis. Given the facts of a new case the system would retrieve an
appropriate 'Thomas' (as it is often referred to in England). As Thomas' analysis
is based on actual cases - which are all referenced in his writings, the next step is
then to directly retrieve those cases that have bearing on the present case, perhaps
even without going explicitly through a 'Thomas'. The contents of the 'Thomas'
can function as conceptual indices for retrieval from the case-base. Some of those
indices are of a special nature and relate to the particular type of crime being
analysed. Others are general, and it is not surprising that we have discovered more
than a partial overlap with the indices we have elicited from our expert.
There are, however, problems with this approach. As already mentioned, this
kind of analysis obviously has no binding legal status (even though it references
real cases), Thomas' work covers only some areas of criminal law, and it is directly
relevant only to England. Additional problems will be discussed in the next section.
[74]
6.2.2. Guideline Sentences

Guideline sentences were pioneered by Lawton L.J. in the 1970s, and later on by
the Lord Chief Justice, Lord Lane. Appellate decisions usually deal with the facts
of a particular case only. But a guideline sentence sets out the parameters of offence
for a particular category of crime, and then lays down a range of sentences for
each such parameter. It does this, not by supplying a table of numbers, but in the
textual style usually found in appellate judgments. The intention in delivering such
guideline judgments may be deduced from the now famous saying:
We are not aiming at uniformity of sentence; that would be impossible. We
are aiming at uniformity of approach. (Lord Lane in: (1980) 2 Cr. App. R. (S)
177)
We shall give some examples of general guideline statements, from two appel-
late cases relating to rape:
Rape is always a serious crime. Other than in wholly exceptional circum-
stances, it calls for an Immediate custodial sentence .... A custodial sentence
is necessary for a variety of reasons. First of all to mark the gravity of the
offence. Secondly to emphasise public disapproval. Thirdly to serve as a warn-
ing to others. Fourthly to punish the offender, and last, but by no means least,
to protect women. The length of the sentence will depend on the circum-
stances. That is a trite observation, but these in cases of rape vary widely
from case to case.
Some of the features which may aggravate the crime are as follows. Where
a gun or knife or some other weapon has been used to frighten or injure the
victim. Where the victim sustains serious injury (whether that is mental or
physical). Where violence is used over and above the violence necessarily
involved in the act itself. Where there are threats of a brutal kind. Where
the victim is very young or elderly. Where the offender is in a position of
trust. Where the offender has intruded into a victim's home. Where the victim
has been deprived of her liberty for a period of time. Where the rape - or
succession of rapes - is carried out by a group of men. Where the offender
has committed a series of rapes on different women, or indeed on the same
woman. (Lord Lane in: Robert (1982) 4 Cr. App.R. (5) 8)
For rape committed by an adult without any aggravating or mitigating fea-

tures, a figure of five years should be taken as the starting points in a contested
case. Where a rape is committed by two or more men acting together, or by
a man who has broken into or otherwise gained access to a place where the
victim is living, or by a person who is an a position of responsibility towards
the victim, or by a person who abducts the victim and holds her captive, the
starting point should be eight years (Lord Lane in: Billain (1986) 8 Cr. App.
R. (5) 88)
[75 ]
180 URI J. SCHILD
As handed down by the Appeals Court, such statements were of course binding
upon the lower courts. Given a particular case a judge could locate that case within
the framework established by the guideline sentence. However, the judge would
still have discretion to adapt the sentence to the special features of the case at hand.
There are several criminological and jurisprudential problems with this ap-
proach to sentencing. These problems will of course be directly relevant to any
computer system seeking to incorporate the approach of guideline sentences.
(1) Only little guidance is given to the weighting of the aggravating and mitigating
factors and their combination. That should not be surprising - such guidance
cannot be given in an explicit form. That would reduce the guideline sentences
to sentencing guidelines.
(2) The guidelines are not always numerically consistent (Ashworth 1984, 521-
522).
(3) The guideline sentences only cover some areas of criminal law.
(4) Guideline sentences could perhaps reduce sentencing disparity in each separate
area, but they do not form a general, overall sentencing policy.
The guideline sentences have been welcomed by the English judiciary, but the
extent of their impact on sentencing disparity is not quite clear.
Our remarks concerning support systems using Thomas' analyses are also true
with respect to guideHne sentences (except of course that the latter are of a legally
binding status). The guideline sentences may perhaps be used to extract rules for
rule-based sentencing systems, but are far more suggestive of the CBR approach,
and even explicitly define some of the indices for retrieval.
6.3. THE CBR PARADIGM A~D SENTENCING
6.3.1. General Overview

Before describing some of the technical aspects of a CBR system for sentencing
support let us attempt to outline the actual use of such a system, and in this way
realize what components the system should be composed of and what capabilities
it should include.
So let us imagine a judge about to pass sentence on an offender and wishing
to use a CBR-system. In order to obtain some guidance he should be able to
characterize the case at hand in some manner, so the system can retrieve a similar
case. The system must of course convince the user that the retrieved case is indeed
'similar' in some definite sense. Using some kind of nearness-measure it should
perhaps even indicate how near the retrieved case is to the case at hand.
The next step for the combined human-computer system is using the sentence
in the retrieved case to arrive at a sentence for the case at hand. The fundamental
assumption is that if the two cases are similar, the retrieved case will supply a
guideline sentence in the sense of Lord Lane's sentences described in Section 6.2.2.
The computer system should present an explicit comparison of the two cases in
[76]
a way that shows mitigating and aggravating aspects of the new case relative to
the old case. The system can then indicate how to adapt the old sentence based
on those aspects, and make appropriate recommendations. The actual adaptation
will of course be done by the judge himself. The system should not recommend
a specific sentence, although that is possible in principle. Some judges would not
appreciate that.
Thus the first item to consider in order to create a CBR-system must be a collec-
tion of criminal cases from which to retrieve a relevant case. The following Section
6.4 will consider various possibilities for this case-base.
In order to enable case-retrieval the cases in the case-base must be indexed in
some manner. The user of the system can then specify values of some of those
indexes for the present case, and retrieve a similar case. Similarity means that the
retrieved case has the same or 'near' values of the selected indexes. Thus the second
point to consider is how to index the case-base. This will be done in Section 6.4.1.
It should be obvious that retrieving just a textual representation of an old case
will not suffice. The comparison of the cases necessitates some structured repre-
sentation of the cases. This structure is considered in Section 6.4.2, and retrieval is
discussed in 6.4.3.
The actual reasoning, i.e., comparison of the cases, weighing the mitigating and
aggravating factors and concluding how to adapt the old sentence to the present
case is the final step, also briefly mentioned in Section 6.4.3.
6.3.2. Our System: Short Example

In order to illustrate the application of a CBR system for sentencing support we
shall consider a concrete example executed on our (third) system. As the system
interface and case-base are in Hebrew, there is no point in showing the actual input
or output. Instead we shall show an edited English version in free format.
Consider the following storl (describing an actual case):
A young man had one night been out driving with a friend of his. The friend had
then suggested they should rob a gas station. The man really had no intention of
getting involved, but his friend eventually made him come along. They were later
apprehended and found guilty. The man had unfolded the entire story in court,
making a rather honest impression.
The system first elicits knowledge about this case by querying the user about the
facts of the case. The word 'facts' is here used in the sense of conceptual indexes,
i.e. nodes in the discrimination trees mentioned in Section 5.2 in connection with
our (second) conceptual retrieval system.
The user must select one or more among the fundamental sentencing principles.
This is one of the retrieval indexes in the case-base. For this case the judge indicated
rehabilitation. As the system is interactive, a judge may experiment with several
9 We shall use the word 'story' as the equivalent of the word 'case'. The motivation is (Schank,
1990).
[77 ]
182 URI J. SCHILD
sentencing approaches, view the suggested sentences of each approach and only
then make up his mind and come to a decision.
Let us assume that for this robbery case the presence of the following facts
(indexes) has been indicated by the user:
(1) use of weapon, (2) no criminal past, (3) additional offenders, (4) offender
is young, (5) amount of money involved is small, (6) offender was influenced by
others, (7) ...
The system will retrieve a case from the case-base relating to the robbery of
a post office by two offenders. The user may retrieve a textual description of this
case, should he wish so. However, the central output is a comparison of the two
cases, i.e. a comparison of the conceptual indexes. This comparison consists of
four parts:
A: Identical indexes: (1) use of weapon, (2) no criminal past, (3) additional offend-
ers, (4) offender is young
B: Similar indexes (i.e. near neighbours in the discrimination tree): none
C: Indexes present in case at hand, but not in precedent: (1) offender was influenced
by others (mitigation), (2) amount of money involved is small (mitigation), (3)
offence committed at night (aggravation), (4) ...
D: Indexes present in precedent, but not in case at hand: (1) amount of money
involved is large (mitigation lO ), (2) part of the money was retrieved (aggravation),
(3) premeditation (mitigation), (4) ...
The system will state that the similarity between the cases is 'high'. Its final out-
put is the recommendation that the sentence in the new case should be 'significantly
lower' than in the retrieved precedent (five years).
The fuzzy values ('high', 'significantly lower', etc.) are precisely defined on
appropriate scales. As explained previously, actual numerical values are not given
unless specifically asked for, as some judges consider the main function of the
system to indicate key characteristics (indexes) of the cases, and compare these. It
is, however possible to request actual numerical results. In this example the system
recommended a sentence of 1 year, while in the real case the judge decided on 13
months.
6.4. SYSTEM ARCHITECTURE

When building a CBR-system for sentencing there are in principle two possibilities
for choosing a case-base. One can make it as large as possible, by including all
10 Mitigation/aggravation in this part of the output means that the new case has a mitigat-
ing/aggravating factor when compared to the precedent.
[78]
available cases. This will facilitate the search for a case as close as possible (in
some sense to be specified) to a new case, but does not necessarily alleviate sen-
tencing disparity. For such a case-base would include cases from different judges,
who execute their individual sentencing policies (and they are not even consistent
with themselves). Given a new case, it may happen that the system retrieves one or
more old cases with very non-characteristic sentences. This approach may even be
worse than the statistical one criticised above.
On the other hand, one may build a case-base selectively, choosing only those
cases whose sentences conform to the current sentencing policy. This would be in
the spirit of Lord Lane's guideline sentences. When passing sentence a judge would
thus not only retrieve past similar cases from his own memory, but the system
would retrieve 'guideline cases' also to be considered. This raises the question
who should make the choice of appropriate 'guideline cases'.
What we therefore propose is the establishment of a select sentencing case-
base by a judicial commission, which has received appropriate authority through
legislation. The case-base should be sufficiently large, so that for any new case
arising it would be possible to find at least one 'near' case in the case-base. In
practice this is of course impossible. There will always be new cases exhibiting
characteristics never considered, or never considered together before. This implies
that the case-base must be continuously added to. It must also be continuously
culled, according to changing views of sentencing policies with time. All these
activities should be the obligation of the judicial commission.
Even if all sentencing judges were to use a CBR support system with this au-
thorised case-base, their sentences would nevertheless not be identical. However,
two judges deciding two similar cases, and applying the same sentencing approach
would receive identical recommendations from the system. More than that one
cannot hope for.
When we decided to develop a CBR prototype for sentencing support, we had
no case-base available, neither a large general one, nor a small select one. We
therefore constructed the case-base from a collection of cases, which we obtained
by interviewing judges. Our object was mainly to explore the way judges thought
about old cases when passing sentence, and consider ways of automating reasoning
using such cases, i.e. build an engine for a CBR system.
It was obvious that a system incorporating such a case-base could not alleviate
sentencing disparity. However, the engine and the case-base are completely sepa-
rated, and technically one can easily substitute another case-base (say a uniform
one, in the same domain) for the present one.
The problems one has to deal with when building a CBR system are:
(1) Indexation: How may cases be retrieved by a case-based reasoner in order
to give advice on quantitative results also for a new case?
(2) Knowledge Representation: How should criminal cases be represented in a
computer.
(3) Reasoning: How should a retrieved case (or cases) be applied to a new case.
[79]
184 URI J. SCHILD
6.4.1. Indexing
The indexing problem raises the question which cases ajudge is reminded of, when
determining the sentence of a new case. A case, usually has one or more special and
outstanding features. In the CBR community such a feature is called an 'anomaly'
of the case (Kolodner 1993, 115), though its appearance may be quite ordinary.
When encountering such an anomaly in a new case, a judge may be reminded about
an old case with the same anomaly. Thus the anomalies could form the indices for
a sentencing case-base.
In order to discover what the anomalies could be, we interviewed judges from
the Tel-Aviv District and Appeals Court. We did not include the judge who had pre-
viously been involved in knowledge elicitation. The interview consisted in asking
the judges to tell us about old cases of theirs, in the two areas of the criminal law we
are concentrating on, Robbery and Rape. It then appeared that when a judge told us
about a case, he would invariably come up with a heading for this case. He would
supply that heading unasked, before or after telling about the case and sometimes
in the middle. This heading turned out to be the most prominent anomaly of the
case.
The case used as an example in the previous section was told by the judge
presiding in the case. The judge explained his way of determining the sentence in
this case, and ended by remarking: "That's what happens when you cannot say no".
It became apparent that the anomalies supplied by the judges were closely corre-
sponding and sometimes even identical to the sentencing parameters (i.e., discrim-
ination tree-nodes) described in the Section 5.2, thus confirming those structures.
These parameters are also equivalent to the ones Thomas and Lord Lane have
described. The judge's remark in the above story we construed to mean, that 'easily
influenced by others' (a node in the offender's discrimination tree from the previous
section) would be an index to the case.
In this case the index indicates mitigating circumstances. Our assumption is that
when the judge mentioned above (and perhaps also other judges familiar with the
story) encounters another case involving an offender with similar characteristics,
he will be reminded of this case. He will also remember the severity of the sentence
- or rather the reasoning behind that sentence as a factor (possibly among others),
that may contribute to the decision in the case at hand.
We thus have sets of almost identical sentencing parameters obtained: (i) By
elicitation from a senior judge. (ii) As headings supplied by other judges. (iii) In
Thomas and Lord Lanes' sentences. (iv) By interviewing jurists and criminologists.
Kolodner discusses the indexing problem, i.e. how to retrieve the right cases at
the right times. She defines two approaches for determining the retrieval indexes:
The functional approach and the reminding approach (Kolodner 1993, p. 203). In
the functional approach one considers the case-domain, the purpose of the retrieval
and how the cases are applied. In the reminding approach one examines the way hu-
[80]
man experts are reminded of cases, in order to determine which kinds of descriptors
are the important ones.
Our activities correspond precisely to these two approaches. Indexing knowl-
edge elicited from the senior judge, extracted from Thomas and Lord Lane and
obtained from expert jurists and criminologists is knowledge determined according
to the functional approach. Indexing knowledge obtained through the interviews
we conducted with the other experts is knowledge determined according to the
reminding approach.
6.4.2. Knowledge Representation

Having determined the indices, the next problem is to create a good representa-
tion of a criminal case. There are several aspects to such a case: What happened,
characteristics and motives of the persons involved, etc.
Consider first the narrative aspects. A 'story' is often considered as consisting
of a sequence of episodes, i.e., events, actions, situations, etc., and the relationships
among such episodes. It is dynamic with little or no hierarchical structure. Classical
AI knowledge structures like semantic networks cannot in themselves suffice for
representing stories. Such basic structures are appropriate for representing certain
static aspects of the stories, but cannot cover the overall picture.
A script (Schank & Abelson, 1977) may actually be an appropriate form for
knowledge representation for the legal process itself, as it may be used to describe
ordinary and routine activities. It is also possible to use scripts to describe crimes,
such as robbery or rape in a generic manner. However, scripts are not appropriate
for describing the reasoning leading to the sentence imposed by a judge.
Also the Memory Organization Packet (MOP) (Schank, 1982; Kolodner, 1983),
which generalises the script describes stereotypical events and does not enable the
kind of explanations we seek.
Narrative understanding systems, e.g., CYRUS (Kolodner, 1981), BORIS (Lehn-
ert et ai., 1983; Dyer, 1983), and MEDIATOR (Kolodner et ai., 1985) are not ap-
plicable here, as our object is not natural-language understanding or story-
understanding, but the application of understood cases to a new case, which needs
to be explained.
One could possibly use various kinds of logics, e.g., Episodic Logic (Schubert
& Hwang, 1989). However, as the originators of such logics usually acknowledge
themselves, much work remains to be done on these logics before they become
applicable in practical systems.
Our decision was to consider the sentencing process as case-based explanation,
and generalise the knowledge structure used for this approach.
6.4.2.1. Explanation Patterns. Swale (Kass et ai., 1986a; Kass, 1986b) is a com-
puter system which produces creative explanations for non-standard stories. Abe
(Kass, 1989; Leake, 1992) is both a simplified and extended version of Swale. The
system defines the concept of an 'explanation pattern' (XP) for a story. It uses the
[81]
186 URI J. SCHILD
explanation patterns for stories in the database to explain a 'gap' in the explanation
of a new story. If these explanations cannot be applied in a straightforward manner,
the system has a number of adaptation strategies.
As a concrete example, assume that the database contains the following two
stories (originally given by Kass, and here considerably edited) and their XPs:
(1) A famous sportsman suddenly collapsed and died. The XP is: Unknown to
everybody he had a weak. heart.
(2) A otherwise healthy lady suddenly died. The XP is: Her husband killed her in
order to obtain the insurance money.
Consider now a new story, based on an actual case: A famous racehorse (called
Swale) suddenly collapsed and died. The 'gap' here is why the horse died suddenly.
The system will adapt and apply the two previous XPs and suggest two possible
explanations for the gap:
(1) Unknown to the owner and trainer the horse had a weak. heart.
(2) The owner killed the horse in order to obtain the insurance money.
We shall call the sentence of a case the 'explanation' of the case. Determin-
ing the sentence of a new case means finding a similar old case and using its
'explanation' (adapting its sentence) to explain the new case.
The use of an XP in connection with a gap can actually work two ways. Given a
story with a gap one can look for stories with an XP to explain the gap. Conversely,
given a gap and its explanation one can look for an appropriate XP in order to
justify this explanation.
A similar situation also occurs in the legal domain. Sometimes a judge will
indeed decide on a sentence after considering the old cases. But sometimes he will
decide on a sentence, and then see how to justify it (both to himself and to the
world), by finding the appropriate precedents. This latter possibility is well-known
and acknowledged by the judges themselves (they sometimes say they have a 'gut-
feeling' of what the sentence should be). Obviously this justification is also an
'explanation' in the ordinary sense of this word.
6.4.2.2. Judicial XPs. An explanation pattern includes the following aspects (slots):
(1) facts, (2) beliefs, (3) goal, (4) action (Schank, 1986; Schank et aI., 1994). We
have adopted this structure, and shall define it precisely below.
When a (decided) case is entered into the case-base, its XP is determined either
by the justification given for written precedents, or by the explanations supplied by
the judge, who told the 'story'. For a new case the facts and beliefs are supplied
by the judge about to pass sentence in the case. He could also supply the purpose,
but should have the capability of experimenting interactively with different crim-
inological approaches. Thus the knowledge structure of both the case at hand and
of cases in the case-base is the XP.
[82]
This approach is somewhat naive. Knowledge about explanation of simple cases

may be represented by a single XP. However, legal cases are complex, and decid-
ing a sentence necessitates the weighing of several factors. The concept of an XP
appears to be insufficient for the kind of explanations we aim at creating. It cannot
cope with the detail and complexity of most legal cases. We have therefore created
a more comprehensive structure, which is made up of a basic MOP and a structure,
called an MXP (Multiple XP), which consists of several XPs. We shall first show
it for the above story.
basic facts MOP: XP-l
accused according to paragraph 402 facts: first offence

found guilty classification: not dangerous to public
baseline sentence: 4 years action: reduce baseline sentence
maximal sentence: 20 years
actual sentence: 1 year
appealed: no
purpose: rehabilitation
XP-2: XP-3:
facts: confessed facts: easily influenced by others

classification: seems trustworthy classification: not dangerous to public
purpose: retribution, rehabilitation purpose: retribution, rehabilitation
action: reduce baseline sentence action: reduce baseline sentence
A Multiple Explanation Pattern (MXP), is a comprehensive structure which is
defined as a collection of viewpoints relating to the sentence in the same crim-
inal case. Each such viewpoint relates to a fact that contributes to (increases or
decreases) the sentence.
Each viewpoint is represented by an XP, and carries a weight relative to the
other viewpoints. It is crucial to understand that an XP as such has no independent
existence. Only the totality of the XPs forming the MXP describes the particular
case and its sentence.
Each XP has the following slots:
1) Facts: This slot contains an index, which is a leaf in the index-hierarchy.
2) Classification: The values of this slot are (intermediate-level) indexes in the
index-hierarchy.
3) Action: The value here indicates the action to be taken with respect to a base-
line sentence. If, e.g., rehabilitation is the sentencing purpose, then the action-
slot could indicate some degree of mitigation.
Two of the basic facts need some further explanation:
(1) Baseline Sentence: This is defined as 'the sentence given for this particular
offence in the absence of any mitigating or aggravating factors'. This definition
[83 ]
188 URI 1. SCHILD
is problematic, as different judges will choose different baselines. Indeed, this is

probably one of the reasons for the existing sentencing disparity.
A judicial commission in Israel headed by a supreme court judge has recently
completed its work and submitted a report to the Minister of Justice. In this report
the majority opinion recommends the establishment of baseline sentences (accord-
ing to the above definition) for all types of crimes in the interest of uniformity.
Should this recommendation be adopted the creation of a uniform case-base will
be easier.
When establishing the case-base the slots and basic facts of the MXP are filled
out for old cases by a legal professional, preferably under the guidance of the sen-
tencing judge himself. The sentence measured out in a case reflects the combina-
tions of XPs in the MXP, though no explicit numerical formula for this combination
is possible.
(2) Purpose: This indicates the sentencing principle applied by the judge (for
each particular case in the case-base), or the sentencing principle the judge wish
to apply (when using the system for a new case). Judges may of course use dif-
ferent approaches in different cases, and they only seldom state explicitly which
sentencing principle they apply in a given case. This information must therefore be
determined a posteriori by the experts analysing the case, perhaps with help from
the sentencing judge.
In connection with the Israeli SIS (Section 3.5) we mentioned that Israeli judges
will be asked to supply this piece of information, which will become part of the
official protocol of the case. However, it is not clear whether the judges will actually
comply with this request.
6.4.3. Retrieval and Reasoning
The retrieval may be implemented in different ways. One could, for example, con-
sider the XPs of an MXP as similar to snippets (see (Kolodner 1988), (Redmond
1990), (Branting 1991)). However, our approach is to select a single, 'best' MXP
using a Tversky contrast-measure, see (Tversky 1977).11
It is not our intention to give detailed descriptions and algorithms for the re-
trieval of old MXPs, choice of the 'best' one and its adaptation to fit the MXP of
a new case, including a proposal of a sentence or a sentencing-range for the new
case. In particular we have avoided any description and discussion of the deter-
mination of numerical weights that necessarily must be appended to the indexes.
The concrete example given in Section 6.3.2 must suffice until detailed publication
elsewhere. We are comparing the CBR paradigm with other paradigms in this
paper, and indeed arguing for it, but not necessarily for the particular approach
we have chosen.
11 For a different approach to similarity of cases for sentencing purposes, see (Murback and Nann
1993).
[84]
6.5. SUMMARY OF CBR APPROACH
We emphasise that our advocacy of this approach is based on the intuitive appeal
this method will have to legal practitioners, and to the ease at which both engine
and case-base may be constructed. We believe the small and select case-base ap-
proach is to be preferred, as it is easily implemented, but also the large case-base
alternative can supply a judge with constructive advice when passing sentence,
assuming of course that a large uniform case-base is available.
One point must still be clarified: Cases in a case-base may appear to be conflict-
ing, just as rules may conflict. There is however a major difference between these
two kinds of inconsistencies. Cases in the case-base reflect real legal situations,
and legal experts are accustomed to resolve conflicts in case-law. They also deal
with inconsistencies in statutory law, but the sentencing rules in a rule-based expert
system are not the law, they have no legal standing. They are superficial creations,
that carry no more weight than the individual judge decides to assign them (which
may be none at all).
It is thus possible and quite acceptable, that the system will create one line of
sentencing recommendation based on one retrieved case, while another retrieved
case will yield another recommendation. The judge will make his own decision
accordingly - this is the essence of human discretion, which no system has.
7. Comparison of Approaches
At the end of each section dealing with a particular kind of support system, we
have summarised our criticism of that particular kind of system. We shall now
discuss some additional features of sentencing support systems, and summarise our
findings. We shall omit any considerations of the American sentencing guidelines
program, as it is of no relevance to our main concerns. Nor shall we deal with
model-based systems, for reasons given in Section 5.3. Thus we shall consider
three kinds of systems:
(i) Statistical Information Systems

(ii) Rule-Based Systems
(iii) Case-Based Systems
We shall consider various aspects of those systems. The financial and budgetary
ones relating to the establishment and maintenance of the systems are of central
importance, but will be considered only in passing. The aspects we shall examine
are:
(1) Maintenance and Scale-up

(2) Complexity and Efficiency
(3) Acceptance by Judiciary
[85 ]
190 URIJ. SCHILD
7.1. MAINTENANCE AND SCALE-UP
These two issues are often neglected in AI work, giving rise to some of the strongest
criticism voiced against AI. Often a prototype system is constructed, but no consid-
eration is given to the feasibility of updating and maintaining its knowledge-base
from time to time. It also happens, that a prototype system is implemented with
a knowledge-base appropriate for a 'toy-world', but the problems arising in an
attempt to apply the system to the real world (scale-up) are not dealt with.
Concerning the transition from a prototype to a real world system there are two
problems to consider.
(A) Is it feasible to obtain all necessary knowledge.
(B) Can the system at all function with a real world knowledge-base.
With respect to maintaining a support system, there are actually three separate
problems to consider.
(C) Who decides when to update the system, and what knowledge to add, delete or
exchange.
(D) How soon will the changes be reflected in the advice given by the system.
(E) How easy is it to carry out such changes.
Only very recently has consideration been given to some of these problems
in the area of AI and Law (Bratley et aI., 1991; Berman & Hafner, 1995; Riss-
land & Friedman, 1995). We shall deal with these aspects for all kinds of sys-
tems mentioned in this paper, notwithstanding our preference for a certain kind of
system.
7.1.1. Statistical Infonnation Systems

(A) Scale-up entails the collection of great numbers of cases from the relevant
jurisdictions. Today this material would be available on magnetic media in many
countries. If case retrieval is to be enabled, much work must be invested in index-
ing and other technical preparations. This is a problem of availability of expert
manpower and sufficient funding.
(B) Problems of scale-up are heavily dependent upon the system architecture.
While the Israeli system is very simple, the approach using existing relational
database environments can lead to unwieldy procedures for maintenance. Never-
theless, even for very large case-bases present-day technologies should enable the
straightforward establishment (and maintenance) of database systems.
The editorial in Crim, L. R., March 1997 directs the attention to another prob-
lem:
The High Court of Justiciary in Scotland has given relatively few authoritative
judgements on sentencing, but in other countries, e.g. in England, the Court
of Appeal's decisions are manifold and ought to be followed. However, some
[86]
judges maintain that there is a big difference between the sentence levels up-
held by the Court of Appeal for certain crimes and sentence levels in practice.
(Crim, L.R. 1997, pp. 153-154).
The New South Wales system considered in Section 3.3 keeps separate data-
bases for sentencing practice and appellate judgements, and each country wishing
to create a statistical sentencing database must decide whether this is indeed the
architecture which is appropriate for the local situation.
(C) Presumably there would be no need for any prior examination of cases to
be added to the system, as the more cases included the better. Thus the process of
maintenance is an automatically on-going process, dependent only on appropriate
funding.
(D) The major problem with statistical systems (in addition to the ones already
detailed in Sections 3.2-3.6) relates to the extent to which the system reflects cur-
rent sentencing policies. The statistical analysis necessarily draws data not only
from very recent decisions, but also from old cases still present in the case-base.
Such cases should be discarded when their sentences are no longer considered
relevant, and somebody has to make that decision. Finally, if too many old cases
are deleted from the case-base, the remainder may no longer form a statistically
significant set.
(E) Much work is needed to review old cases, and it is not clear who should be
responsible for such work.
7.1.2. Rule-Based Systems

(A) Scale-up of a rule-based system entails establishing rule-bases for all types
of offence. Even if one were to imagine that agreement can be reached among
the judges about the formulation of such rules with respect to a single type of
offence, one can hardly expect agreement about general rules relating to the entire
sentencing domain.
(B) Assume that agreement can be reached, and that the rule-base is properly
constructed. This means that the rule-based system exhibits a strict division be-
tween the engine and the knowledge-base. Not all rule-based systems observe this
feature, e.g. the Israeli system for probation officers described in Section 4.1.1.
In the former case, scale-up will be additive (and hence relatively easy) in the
following sense: A rule-base for the entire criminal domain will consist of a general
part and the union of smaller rule-bases relating to each individual type of crime.
Thus going from a system addressing only selected criminal offences to a larger
system addressing additional offences, will imply only the addition of rules, and
no change in already existing rules. With proper annotation of the rules such a
system would not exhibit special technical problems. In the latter case, scale-up as
such may be impossible.
(C) We have previously (Section 4.1.4) expressed the view that no initial rule-
base can be constructed, which all the judges would agree upon. Such disagreement
[87 ]
192 URI J. SCHILD
would of course also carry over to the problem of maintenance: Who shall decide
which rules to add, delete or change.
(D) If agreement on maintenance is reached, changes in the knowledge-base
would of course have immediate effect on the advice of the system.
(E) If a division exists between-engine and rule-base and the rules are properly
annotated, maintenance can be carried out. If not, updating the rules may be a
technically difficult problem. For a very large system it may even be an impossible
task. There is evidence that rule-based systems with more than a few thousand rules
cannot be properly maintained.
7.1.3. Case-Based Systems
In Section 6.4 we distinguished between two types of CBR systems for sentencing.
One of those would have as large a case-base as possible, while the other one
would consist of a much smaller collection of select landmark cases. The points
we discuss must be considered separately for each of those possibilities.
(i) Large Case-Base

(A) As we have explained above for statistical information systems, the collection
and preparation of cases is a technical and financial problem. Indexing is a problem
very relevant to the scale-up from a small prototype of a CBR system to a system
with a large case-base. Automatic indexing is not feasible at present, though much
research is being done on that topic using methods of Natural Language Processing
or other statistical approaches (Chen, 1993; Ginsberg, 1993; Schweighofer, 1995).
We are therefore dealing with manual work, which has to be carried out by experts.
(B) Whether a case-based reasoner would be able to function with a very large
case-base remains to be seen. Most CBR systems built so far have been prototypes
only, and it is difficult to foresee which problems may arise ifthe case-base includes
tens of thousands of cases.
(C) The extremely big problem we described for statistical information systems
also exists here. Who shall decide which cases to delete, and who will have the
expertise, time, etc. to carry out a constant review of a large case-base.
(D) Changes in sentencing policy will be reflected the moment new up-to-date
cases are added, and old, no longer relevant cases are deleted.
(E) The actual changes are at most a technical problem. However, this problem
may be severe, if nodes of the index-trees have to be changed. Such a change will
imply a similar change in the indexing structure of the individual cases of the case-
base. This is no easy task, which at present cannot be carried out by an automatic
procedure.
Changes in the index-tree structure would not be a frequent occurrence. How-
ever, changes in the weighing function (recalibration) could sometimes be neces-
sary. This problem is technicaL and not too severe.
[88]
(ii) Select Case-Base

There are two possibilities relating to which cases should be included in a select
case-base for sentencing.
(1) A judge may be willing to include only his own previous sentencing decisions
in what will then become his own private case-base. It would not contribute to
the aim of attaining general uniformity in sentencing, but at least the judge's
own sentences may become uniform.
(2) The establishment of a public panel, perhaps consisting of judges, practising
lawyers, academics and legislators. This panel would decide which new cases
to add and which old cases to delete from the case-base, thus ensuring that the
case-base will reflect the sentencing policies in a faithful manner. It is an open
question whether such a panel could reach an agreement about the inclusion of
cases.
(A) Without any experience at present, it is impossible to say how a public commit-
tee would reach agreement on which cases to include in the select case-base.
Obviously a lot of politics would be involved. Technically, however, there are
no problems.
(B) CBR systems exist which function well with small case-bases.
(C) As already explained a public committee will make the decisions. The actual
update is technically straightforward
(D) Changes in sentencing policy will be felt immediately with the appropriate
changes in the case-base.
(E) Updates are easily carried out. Even changes in the index-tree structure and
appropriate changes in the indexing of the cases in the case-base would not be
extensive, as the case-base is assumed small.
7.2. COMPUTATIONAL COMPLEXITY AND EFFICIENCY

7.2.1. Classification of Problems
The efficiency of a sentencing support system may be considered from three points
of view.
(A) The effect of a system on the expenses of the justice system. It is a well-
known feature of the courts of law in many countries that they operate at great
cost and usually with a great backlog of cases. It is possible in principle that the
introduction of computer systems of the type described above may reduce those
costs.
Consider for example an expert system, which identifies and recommends bail
decisions for individual cases. A prototype for such a system is described in (Has-
sett, 1994), and was discussed in Section 4.1.2. The financial burden imposed on
the state by pre-trial detention is considerable. A computer system could perhaps
identify the persons not to be detained in a more efficient manner than today, and
thus lower the costs
[89 ]
194 URI J. SCHILD
However, it is far from clear whether this indeed will be the case. Experience
shows that the introduction of computer systems usually increases the amount of
information available to human users of the system, but does not decrease the
costs. There are also many examples of environments where throughput was not
increased through computerisation.
We therefore do not believe any conclusions can be drawn about the three types
of sentencing support systems with respect to costs and expenses, especially as the
costs of establishing and maintaining the system also have to be taken into account.
(B) What is the utility of the system? i.e., does use of the system actually im-
prove the work of the judge using it? Does it raise the quality of his decisions? To
establish criteria and benchmarks for the utility of intelligent computer systems in
the legal domain is a general problem, that has not been dealt with so far. At present
the only possible method to examine this issue must be by interviewing the users.
No such data is available for any system mentioned in this paper.
(C) The purely computational aspect, i.e., whether the system interacts suffi-
ciently fast with the user. As no data is available concerning the first two issues, we
shall only deal with the third one for the three kinds of system we discuss.
7.2.2. Comparison of Systems

Statistical Information Systems:
In the case of the Scottish prototype the developers admit to severe limitations
of resources [Hutton et al. 1995,94], even though the case-base is quite small (4500
cases). We are concerned this will always be the case, when database technology
is applied to what is essentially a straight-forward statistical problem. One of the
main justifications of the Israeli approach to system architecture is the capability
of attaining computational efficiency even with a large sentencing base.
Rule-Based Systems:
The efficiency of even very large rule-based systems has been well established.
There is no reason why a rule-based system in the sentencing domain should not
be efficient, once the rule-base has been created, especially if the rule-base is
partitioned according to the crimes.
Case-Based Systems:
A CBR system with a small and select case-base will clearly be fast, even if
its engine is not computationally efficient. For systems with a large case-base no
previous experience is available.
7.3. ACCEPTANCE BY THE JUDICIARY

We believe that the major principle of guidance in constructing a sentencing sup-
port system should be its ultimate acceptance and actual use by the judges. A
prototype will of course always be of theoretical interest. But there would be no
[90]
real purpose in fully developing such a computer system, if it would be applied by

just a very small number of judges, used only by researchers, or not used at all.
It may not be easy to gain general acceptance for a decision support system by
the judiciary, though the use of computers is a common phenomenon in a court
today. The principal use of computers is for word-processing and sometimes for
textual information retrieval, using a system like LEXIS or Westlaw. The acquisi-
tion of an expert advisory system should therefore not be problematic, for it does
not introduce completely new technologies. .
However, a system for sentencing support may be met with (perhaps passive)
resistance for quite different reasons. One of the judges we interviewed in the
project described above, told how both supplying and receiving sentencing in-
formation have been met with opposition by many judges. As president of the
Tel-Aviv District and Appeals Court she had requested the judges of the court to
send her short monthly reports about their criminal cases, stating briefly the facts
and the sentences. These reports were to be distributed among all the judges of the
court, with the intention of gaining if not uniformity, then at least a common and
up-to-date basis for sentencing. Only few judges complied with her request, several
told her they were against receiving (not to mention supplying) such information,
and after some months she decided to stop the experiment.
When asked the reasons for the refusal of the judges, she indicated that perhaps
these judges did not believe brief summaries were of any use, and that nobody had
the time to write (or read) long protocols of other judges.
There appears to be a basic inconsistency here. We have already pointed out
(in Section 2.1) that sentences should be analysed with respect to the motivation
of the judge. Each individual judge is influenced and reacts differently to the im-
pression made by the offender, the witnesses, etc. in court. Such influences are
often very subtle, and can never be apparent from the records. Even the inclusion
of a sentencing approach parameter is only a crude summary of this non-explicit
information.This is of course well-known to the judges themselves, and may be
the reason for the rejection by some of them of only short summaries. Such judges
might presumably not want to know what a computer system has to say about the
sentencing possibilities in a given case.
On the other hand, all the judges we interviewed agreed that statistical sentenc-
ing data is 'a good thing' and should be made available, even if based on only
a small number of parameters. The initiative to add information on sentencing
approach comes from our criminologists, and lack of this parameter did not seem
to bother the judges.
It should be obvious that no judge can ever be forced to use such a system,
and more subtle ways would be needed in order to convince the judiciary to use
a decision support system. At this point let us just remark, that as the acceptance
of this kind of system is problematic for external (i.e. criminological and jurispru-
dential) reasons, a software developer must take good care that it should at least be
[91]
196 URI J. SCHILD
extremely user-friendly and supply the user with sufficient information so that he
should not also have to apply manual labour for searching precedents, etc.
7.3.1. Statistical InfO/mation Systems

We have already explained above that the judges seem to want precisely this kind
of system, despite the lack of confidence in the data. The problem here seems to be
that this kind of system could be too warmly accepted, in the sense that too much
trust could be uncritically placed in its data.
7.3.2. Rule-Based Systems

It is difficult to imagine that rule-based systems can be accepted by the profes-
sional jUdiciary. There is of course always a possibility to impose a set of rules
by legislation. However, the system thus obtained would be equivalent to a sen-
tence guidelines system (Section 3.1), an approach not relevant to our present
considerations.
7.3.3. Case-Based Systems

When approaching the problem of building a case-based sentencing support system
in Israel or in most other countries outside England, one has neither Thomas nor
guideline sentences to build on. There are, however, a large number of reasons for
using the CBR approach. Much of what is usually said in favour of this paradigm
does not amount to more than hype, in the sense that it may seem intuitively correct,
but has not been backed up by real proof. Nevertheless, in the legal domain it is
particularly convincing, as the use of previous cases is a central aspect of legal
reasoning in general, and in the area of sentencing in particular. We shall therefore
make the following observations:
(1) Judges are accustomed to work with cases, to apply them and to distinguish
them. A case-based advisory system presents the judges with real cases and
sentences, not made-up rules. If judges are at all ready to use a computer for
sentencing support, cases is the natural medium for conveying information.
The reason for this is that cases represent experienced situations. When a similar
situation arises, those decisions and the knowledge that went into making them
provide a starting point for solving the problem the new case poses. In other words,
using the CBR approach it will be possible for a judge to determine a sentence
based on general standards but also to consider the individual circumstances of the
offender and the case at hand.
(2) When a judge has to pass sentence in a new case, he often considers old and
similar cases and may adapt an old sentence to fit the new case. These old cases
are not only the (perhaps binding) precedents he must consider, but primarily
cases from his own experience, or cases his colleagues have told him about.
[92 ]
A judge we interviewed told us how he actually keeps a card-index of his old

cases and their sentences, which he flips through when he has to pass sentence
in a new case.
Judges may not always be reminded of the most appropriate set of cases when
they reason. There is evidence, that when there is much to remember, people cannot
always access the right cases when they need them (Brooks et al., 1989; Gentner,
1989). A CBR system for sentencing will make the appropriate cases available,
thus acting as an auxiliary memory for the judge. It will also help toward an
analysis, but of course leave the final decision to the user.
Using a CBR system with a case-base made up of the judge's own cases would
not be so very different from using a card-index. Of course, the only kind of consis-
tency and uniformity to be obtained would be of the judge with himself. However,
a system with this kind of private case-base could be a transitional step. It would
accustom the judiciary to use a CBR system, where the case-base eventually would
be made up by a public committee.
8. Conclusion
We have presented and analysed some approaches to advisory computer systems

for criminal sentencing. From the jurisprudential and criminological background
it should be apparent that there are many problems that arise when attempting to
create such a system, but no ideal solution. In fact, most problems are not compu-
tational and do not arise from the limitations of computers or even from the area of
Artificial Intelligence. They relate to inherent problems of the domain.
It is our opinion that among these approaches the case-based one is the most
practical and feasible one. It is not perfect, but it seems optimal. This takes into
account the objective problems of other kinds of systems, and the opposition from
the judiciary itself to these other kinds of Systems. A CBR system is a tool which
will not be too unfamiliar to legal experts, who anyway work with cases, and it is
relatively easy to build.
Two major questions arise with this kind of system:
1) Can agreement be reached about which cases to include in the case-base?

2) Will the judges be ready to devote sufficient time to read and consider the
analysis and arguments of the system?
It is our hope that positive answers may be found to both questions. The intro-
duction of our present prototype on a limited trial basis in the Israeli courts will
test this hypothesis.
[93 ]
198 URI J. SCHILD
Acknowledgements
The programs at Bar-Ilan University were written by Gil Ehrlich and Yaakov Kerner
respectively. I thank the judges of the Tel-Aviv District and Appeals Court and
especially its vice-president, Judge M. Talgam for extensive support.
References
Alty,. J.L. and Coombs, MJ. 1984. Expert systems. Manchester: NCC Publications.
Ashley, KD. 1990. Modelling Legal Argument. Cambridge: MIT Press.
Ashley, K.D. 1992. Case-based reasoning and its implications for legal expert systems, Artificial
Intelligence and Law I, 113-208.
Ashworth, A. 1984. Techniques of guidance on sentencing, The Criminal Law Review 519-530.
Ashworth, A. 1991. Sentence reform structures, In M. Tonry (ed.), Crime and Justice - A Review of
Research. Chicago: Chicago University Press, 16: 181-241.
Ashworth, A. 1992. Three techniques for reducing sentence disparity. In A. von Hirsch and A.
Ashworth (eds.), Principled sentencing. Edinburgh: Edinburgh University Press, pp. 282-290.
Ashworth, A. 1995. Sentencing and Criminal Justice. 2nd edn., London: Butterworth.
Bain, W.M. 1986. Case-based reasoning: A computer model of subjective assessment. Ph.D. thesis.
Yale University.
Berman, D.H. and Hafner, C.D. 1989. The potential of artificial intelligence to help solve the crisis
in our legal system, Communications of the ACM 32,928-938.
Berman, D.H. and Hafner, C.D. 1995. Understanding precedents in a temporal context of evolving
legal doctrine. In Fifth international Conference on Artificial Intelligence and Law (ICAIL-95).
College Park, MD: ACM Press, pp. 42-51.
Bing, J. 1987. Designing text retrieval systems for conceptual searching. In Proceedings of the First
International Conference on Artficialintelligence and Law (ICAIL-87). Boston: ACM Press, pp.
43-51.
Bing, J. 1989. The law of the books and the law of the files. In G.P.v. Vandenberghe (ed.), Advanced
topics of law and information technology. Amsterdam: KIuwer, pp. 151-182.
Branting, L.K 1991. Reasoning with portions of precedents. In Proceedings of the Third Inter-
national Coriference on Artificial Intelligence and Law (ICAIL-91). Oxford: ACM Press, pp.
145-154.
Bradey, P., Fremont, J., Mackaay, E., and Poulin, D. 1991. Coping with change. In Proceedings of
the Third International Conference on Artificial Intelligence and Law (ICAIL-91). Oxford: ACM
Press, pp. 69-76.
Brooks, L.R., Allen, S.w., and Norman, G. 1989. The multiple and variable availability of familiar
cases. In C. Hammond (ed.), Proceedings: Workshop on case-based reasoning (DARPA). San
Mateo, California: Morgan Kaufmann, pp. 130-132.
Burrows, J., Henderson, P., and Morgan, P. 1994. Methodology: A project to improve bail decisions.
Presented at: Conference of the American Society of Criminology. Miami, USA.
Chan, J. 1991. A computerized sentencing information system for New South Wales courts,
Computer Law and Practice 137-150.
Chandrasekaran, B. and Mittal, S. 1982. Deep versus compiled knowledge approaches to diagnostic
problem-solving, Proceedings AAAI-S2. Los Altos, California: Morgan Kaufman Publishers, pp.
349-354.
Chen, H., Lynch, KJ., Basu. K., and Ng, T.D. 1993. Generating, integrating and activating thesauri
for concept-based document retrieval, IEEE Expert 8(2), 25-34.
Clancey, W.J. 1983. The epistemology of a rule-based expert system - a framework for explanation,
Artificial intelligence 20 215-251.
[94]
Davis, R., Buchanan, B., and Shortliffe, E. 1977. Production rules as a representation for a
knowledge-based consultation program. Artificial Intelligence 8, 15-45.
Davis, R. and Lenat, D.B. 1982. Knowledge-based systems in artificial intelligence. New York:
McGraw -Hill.
DeMulder, R.Y. and Gubby, H.M. 1983. Legal decision making by computer: An experiment with
sentencing, Computer/Law lournal IV, 243-303
Denning Lord Alfred Thompson 1953. Royal Commission on Capital Punishment, 1949-53 Report,
Cmd. 8932, London: Her Majesty's Stationery Office.
Dewey, J. 1924. Logical method and law, Cornell Law Quarterly to, 17-27.
Dick, J.P. Representation of legal text for conceptual retrieval. In Proceedings of the Third Inter-
national Conference on Artificial Intelligence and Law (ICAIL-91), Oxford: ACM Press, pp.
244-253.
Doob, AN. and Park, N.W. 1987. Computerized sentencing information for judges, Criminal Law
Quarterly 30, 54-72.
Dreyfuss, HL 1992. What Computers Still Can't Do. Cambridge, Mass.: MIT Press.
Dyer, M.G. 1983. In-depth understanding. Cambridge, Mass.: MIT Press.
Forbus, K.D. 1988. Quantitative physics: Past, present and future. In H. Shrobe (ed.), Exploring
artificial intelligence. San Mateo, California: Morgan Kaufmann.
Fox, J.A 1981. Models in quantitative criminology. New York: Academic Press.
Freed, S.J. 1992. Federal sentencing in the wake of guidelines: unacceptable limits of the discretion
of sentences. Yale Law lournal101(8): l681-l754.
Gardner, A. 1987. An artificial intelligence approach to legal reasoning. Cambridge, Mass.:
MITlBradford Books.
Garland, D. 1991. Sociological perspectives in punishment. In M. Tonry (ed.), Crime and justice - a
review of research, Vol. 14. Chicago: University of Chicago Press, pp. 115-165.
Gentner, D. 1989. Finding the needle: Accessing and reasoning from prior cases. In C. Hammond
(ed.), Proceedings: Workshop on case-based reasoning (DARPA). San Mateo, California: Morgan
Kaufmann, pp. 137-143.
Ginsberg, A 1993. A unified approach to automatic indexing and information retrieval, IEEE Expert
8(5), 46-56.
Greenleaf, G. 1991. Information technology and the law, Australian Law lournal65: 45-48.
Hafner, C.D. 1978. An infonnation retrieval system based on a conceptual model of legal knowledge.
Ph.D. Thesis. Ann Arbor, Michigan: University of Michigan.
Hafner, C.D. 1987. Conceptual organization of case law knowledge bases. In First International
Conference on Artificial Intelligence and Law (ICAIL-87). Boston: ACM Press, pp. 35-42.
Hart, H.AL. 1963. Law liberty and morality. Oxford: Oxford University Press.
Hassett, P. 1993. Can expert system technology contribute to improved bail conditions, Int. l. of Law
and Infonnation Technology 1, 144-188.
Hassett, P. 1994. Private communication.
Hogarth, J. 1988. Computer and the law: Sentencing database system, user's guide. Vancouver: LIST
Corporation.
Home Office White Paper 1996. Protecting the public: The government's strategy on crime in
England and Wales. London: Her Majesty's Stationery Office.
Hudson, B. 1987. Justice through punishment - A critique of the justice' model of correction.
London: Macmillan.
Hutton, N., Patterson, A, Tata, C., and Wilson, J. 1995. Decision support for sentencing in a common
law jurisdiction, Fifth International Conference on Artificial Intelligence and Law (ICAIL-95).
Washington D.C.: ACM Press, pp. 89-95.
Johnson, L. and Keravnou, E.T. 1985. Expert systems technology. London: Abacus Press.
[95 ]
200 URI J. SCHILD
Kass, A.M., Leake, D.B., and Owens, C.C. 1986a. Swale: A program that explains. In R.C. Schank
(ed.), Explanation patterns: Understanding mechanically and creatively. Hillsdale, New Jersey:
Lawrence Erlbaum Associates, pp. 232-254.
Kass, A.M. 1986b. Modifying explanations to understand stories, Proceedings of the Eighth Annual
Conference of the Cognitive Science Society. Amherst, MA.
Kass, A.M. 1989. Adaption·based explanation: Extending script/frame theory to handle novel input,
Proceedings of IlCA/-89. San Mateo, California: Morgan Kaufmann, pp. 141-147.
Kass A.M. 1990. Developing creative hypotheses. Technical Report #6. Northwestern University,
Illinois: Institute for the Learning Sciences.
Kolodner, J.L. 1981. Organization and retrieval in a conceptual memory for events, Proceedings of
IlCA/-81. San Mateo, California: Morgan Kaufmann, pp. 227-233.
Kolodner, J. 1983. Maintaining organization in a conceptual memory for events, Cognitive Science
7: 281-328.
Kolodner, J.L., Simpson, R.L., and Sycara-Cyranski, K. 1985. A computer model of case-based rea-
soning in problem solving. Proceedings of IlCAI-85. San Mateo, California: Morgan Kaufmann,
pp. 284-290.
Kolodner, J.L. 1988. Retrieving events from a case memory: A parallel implementation, Proceedings
of a Workshop on CBR. San Mateo, California: Morgan Kaufmann, pp. 233-249.
Kolodner, J. 1993. Case-based reasoning. San Mateo, California: Morgan Kaufmann.
Leake, D. 1992. Evaluating explanations: A content theory, Hillsdale, New Jersey: Lawrence
Erlbaum Associates.
Lehnert, W, Dyer, MG., Johnson, P, Yang, c., and Harley, S. 1983. BORIS - An experiment in
in-depth understanding of narratives, Artificial Intelligence 20, 15-62.
Lovegrove, A. 1987. An evaluation of judicial models for sentencing guidelines In D.C. Penning-
ton and S. Lloyd-Bostock (eds.), The psychology of sentencing: Approaches to consistency and
disparity. Oxford: Oxford University Press, pp. 133-142.
Lovegrove, A. 1989. Judicial decision making sentencing policy and numerical guidance. Berlin:
Springer Verlag.
Martinson, R. 1974. What work? - Questions and answers about prison reform', The Public Interest
35,22-54.
Mital, v., Stylianou, A., and Johnson, L. 1991. Conceptual information retrieval in litigation support
systems. In: Proceedings of the Third International Conference on Artificial Intelligence and Law
(ICAIL-91). Oxford: ACM Press, pp. 235-243.
Murbach, R. and Nonn, E. 1993. Similarity in harder cases. Proceedings of the Fourth International
Conference on Artificial Intelligence and Law (ICAJL-93). Amsterdam: ACM Press, pp. 236-
244.
Posner, R. 1985. An economic theory of criminal law, Columbia Law Review 85, 1193-1231.
Redmond, M. 1990. Distributed cases for case-based reasoning; facilitating use of multiple cases,
Proceedings of AAAI-90. Boston: American Association for Artificial Intelligence, pp. 304-309.
Reynolds, D. and Beck, T. 1993. Tennessee offender management information system, AI Magazine
14(3): 61-68.
Riesbeck, C.K. and Schank, R.c. 1989. Inside case-based reasoning. Hillsdale, New Jersey:
Lawrence Erlbaum Associates.
Rissland, E.L. and Ashley, K.D. 1987. A case-based system for trade secrets law, First International
Conference on Artificial Intelligence and Law. Boston: ACM Press, pp. 60-66.
Rissland, E.L. and Friedman, M.T. 1995. Detecting change in legal concepts. In: Fifth International
Conference on Artificial Intelligence and Law (ICA/L-95). College Park, MD: ACM Press, pp.
127-136.
Salton, G. 1989. Automatic Text Processing. Reading, Mass.: Addison-Wesley.
Schank, R.c. and Abelson, R.P. 1977. Scripts, plans, goals and understanding: An inquiry into
human knowledge structures. Hillsdale, New Jersey: Lawrence Erlbaum Associates.
[96]
Schank, RC. 1980. Language and memory, Cognitive Science 4(3): 243-284.
Schank, RC. 1982. Dynamic memory. Cambridge: Cambridge University Press.
Schank, R.C. 1986. Explanation patterns. Hillsdale, New Jersey: Lawrence Erlbaum Associates.
Schank, RC. and Leake, D.B. 1989. Creativity and learning in a case-based explainer, Artificial
Intelligence 40: 353-385.
Schank, RC. 1990. Tell me a story - a new look at real and artificial memory. New York: Charles
Scribner's Sons, Macmillan Publishing Company.
Schank, RC., Kass, A., and Riesbeck, c.K. (eds.) 1994. Inside case-based explanation, Hillsdale,
New Jersey: Lawrence Erlbaum Associates.
Schild, U.J. 1992. Expert Systems and Case Law. Chichester: Ellis Horwood.
Schubert, L.K. and Hwang, C.E. 1989. An episodic knowledge representation for narrative texts, First
International Conference on Principles of Knowledge Representation and Reasoning. Toronto,
Canada, pp. 444--458.
Schweighofer, E., Winiwarter, w., and Merkl, D. 1995. Information filtering: The computation
of similarities in large corpora of legal texts. In: Fifth International Conference on Artificial
Intelligence and Law (ICAIL-95). College Park, MD: ACM Press, pp. 119-126.
Sergot, M.J. 1991. The representation of law in computer programs. In T. Bench-Capon (ed.),
Knowledge-based systems and legal applications. London: Academic Press, pp. 3-67.
Shapira, M. 1990. Computerized decision technology in social service, International Journal of
Sociology and Social Policy 10, 138-164.
Simon E. and Gaes, G. 1989. ASSYST - computer support for guideline sentencing, Second Inter-
national Conference on Artificial Intelligence and Law (ICAIL-89). Vancouver: ACM Press, pp.
195-200.
Skalak, D.B. and Rissland, E.L. 1992. Arguments and Cases, Artificial Intelligence and Law 1, 3--44.
Smith, J.C., Gelbart, D., MacCrimmon, K., Atherton, B., McClean, J., Shinehoft, M., and Quintana,
L. 1995. Artificial intelligence and legal discourse: The ftexlaw legal text management system,
Artificial Intelligence and Law 3(1-2): 55-95.
Sowa, J.P. 1984. Conceptual structures: Information processing in mind and machine. Reading,
Mass.: Addison-Wesley.
Susskind, R. 1987. Expert systems in law. Oxford: Oxford University Press.
Thomas, D.A. 1979. Principles of sentencing. London: Heinemann, 2nd ed.
Thomas, D.A. 1982. Current sentencing practice. London: Sweet and Maxwell.
Tonry, M. 1993. Sentencing Commissions and their Guidelines. In M. Tonry (ed.), Crime and Justice
- A Review of Research. Chicago: Chicago University Press, 17: 137-195.
Tonry, M. 1993a. The failure of the U.S. Sentencing Commission's guidelines, Crime and Delin-
quency 131.
Toulmin, S.E. 1958. The uses of argument. Cambridge: Cambridge University Press.
Turtle, H. 1995. Text retrieval in the legal world, Artificial Intelligence and Law 3(1-2),5-54.
Tversky, A. 1977. Features of similarity, Psychological Review 84(4), 327-352.
van Noortwijk, c., Piepers, P.A.W., van der Wees, J.G.L., and De Mulder, R.Y. 1991. The JURICAS
system: New applications and future developments. In Proceedings of the Third International
Conference on Artificial Intelligence and Law (ICAIL-91). Oxford: ACM Press, pp. 201-206.
von Hirsch, A. 1992. The Swedish Sentencing Law. In A. von Hirsch and A. Ashworth (eds.),
Principled Sentencing, Boston: Northeastern University Press, pp. 292-307.
von Hirsch, A. and Ashworth, A. 1992. Principled Sentencing. Boston: Northeastern University
Press.
von Hirsch, A. 1993. Censure and Sanctions. Oxford: Oxford University Press.
Weatherburn, D., Crettenden, I., Bray, R, and Poletti, P. 1988. New South Wales sentencing
information system. New South Wales: Judicial Commission of New South Wales.
Welibank, M. 1983. A review of knowledge acquisition techniques for expert systems. Ipswich: British
Telecom.
[97 ]
202 URlJ. SCHILD
Wilkins, L.T., Kress, J.M .. Gottfredson, D.M., Calpin, J.C., and Gelman, A.M. 1978. Sentenc-
ing guidelines: Structuring judicial discretion. Washington D.C.: United States Department of
Justice.
Wilson, J. 1996. Personal communication.
Zeleznikow, J. and Hunter, D. 1994. Building Intelligent Legal Infonnation Systems. Deventer, the
Netherlands: Computer/Law Series, KIuwer Law and Taxation Publishers.
[98]
Artijiciaiintelligence and Law 6: 203-230, 1998.
203
The Application of Judicial Intelligence and 'Rules'

to Systems Supporting Discretionary Judicial
Decision-Making *
CYRUSTATA
Centre for Sentencing Research, Law School, University of Strathclyde, Glasgow, Scotland,
G40RQ.
E-mail: [email protected]. uk
Abstract. This article critically examines approaches to the production of systems of support for
discretionary judicial decision-making in sentencing. The aim of the article is to attempt to illuminate
the character of discretionary judicial decision-making and how academic research has informed the
attempt to model the exercise of judicial discretion. Briefly placing the development of decision
support systems for judicial sentencing in the context of world-wide themes in sentencing reform,
the article proceeds to focus on various attempts to produce systems of decision support for sentenc-
ing. It then briefly explores two of the key issues ('impact' and 'institutional authority') which may
determine the future support of such systems. If systems of decision support for judges are to have
a future then they must not only be accepted at an institutional level, but also be seen as valuable
by those for whom they are designed. Thus, the application of 'judicial intelligence' is unavoidable.
Underlying the judicial sentencing decision process is some conception of 'similarity' between cases.
How, then, should this 'similarity' be represented? Traditionally, representations of similarity have
tended to be informed by 'the legal-analytical' paradigm. The supposedly basic building blocks of
case information ('offence' and 'offender' and 'aggravating and mitigating' factors) are critically
considered. It is argued that systems based on a 'legal-analytical' paradigm are limited in their rep-
resentation of the decision process. These limitations, it is suggested, may be overcome by adopting
an approach which tries to represent the informal schema of understanding which decision-makers
employ and the holistic way in which they think about a case. It is argued, therefore, that judicial
decision-making is amenable to modelling through the use of computer technology, but that there is
a need to re-model our conception of judicial 'intelligence' on which such technology relies.
1. Introduction
Is it possible to build a system of computer-support to aid discretionary judicial
decision-making? If so, how can that decision process be understood and that
understanding be most appropriately represented? This article discusses the appli-
* This article has been very substantially developed from a paper published in the Jour-
nal for Law, Information and Technology: Tata, C, Wilson, J.N.& Hutton, N (1996) 'Rep-
resentations of Knowledge and Discretionary Decision-Making by Decision-Support Systems:
the Case of Judicial Sentencing', 2. The Journal of Information, Law and Technology (JILT).
<http://elj.warwick.ac. ukleljljiltlartifintl2tatal>
[99]
204 CYRUSTATA
cation of different conceptual approaches to information technology to the judicial

sentencing process. Perhaps more than any other group of users in legal domains,
the effectiveness of a decision support system for judges is determined by judicial
acceptance. While in other areas, users may feel obliged to consult a decision
support system, in sentencing, judges who may be confronted with a system which
seems alien to judicial intuition can immediately invoke the powerful currency
of 'judicial independence' to eliminate the expectation that judges ought to be
influenced by a such computer system, or, indeed any other form of systematic
information or education (see for example, Armytage, 1995).1 Thus judges enjoy
wide legally-defined 'discretion' in deciding whether or not to accept the introduc-
tion of a decision support system. How can decision support systems be accepted
by judges and be genuinely effective? This article examines how the conceptual
content of computer programmes intended to support judicial decision-making can
increase the likelihood of their acceptance and usefulness to their users.
Consideration of judicial sentencing support systems as a means of allowing
reform of sentencing practice necessitates at least some discussion of reforms
world-wide in recent years. Over the last twenty years, many Western jurisdictions
have taken various steps to reform the sentencing process (Ashworth, 1992b; Tata,
Hutton, Wilson, & Paterson, 1996; Hutton & Tata, 1995).
2. International Reforms in Sentencing

The primary aim of this wave of international reforms has been to reduce dis-
parity and promote consistency in sentencing (Ashworth, 1992a). Consistency in
sentencing requires that like cases are treated in a similar way and conversely
that dissimilar cases receive different sentences (Council of Europe, 1993; Hutton,
Paterson, Tata, & Wilson, 1996).
It has been the 'just deserts' approach to sentencing which has, in part, in-
fluenced most of the approaches to sentencing reform (Ashworth, 1992a). This
approach argues that sentence ought to be proportionate to the seriousness of the
offence and not based on the character or past conduct of the offender (von Hirsch,
1976, 1993). Reforms in Scandinavia have used a narrative form of guidelines
to guide sentencers as to how the principle of desert should be properly applied
(von Hirsch & Jareborg 1989; Jareborg, 1994). Arguably, the Criminal Justice
Act (1991) adopted a similar approach for England and Wales. In general, the US
approach to sentencing reform has been to construct numerical guidelines which
specify a limited range of penalty for particular offence categories (Wilkins et aI.,
1 It is, of course, arguable that sentencing does not necessarily have to be located at ajudiciallevel.
Location of sentencing powers within Executive institutions would not necessarily offend classical
scholarly theories of 'judicial independence' (indeed many sentencing or quasi-sentencing functions
have been acquired by the Executive, such as the prosecutor fine, parole and remission etc.). However,
my point here is to stress the power of assumed 'commonsense' notions of 'judicial independence'
which proclaim judicial ownership of sentencing.
[100 ]
THE APPLICATION OF JUDICIAL INTELLIGENCE 205
1978). These were introduced into a number of state jurisdictions, the most well
known being the Minnesota Guidelines (Tonry, 1996), and most controversailly, in
1987 the U.S.-wide Federal Guidelines were introduced. The relative inflexibility
of numerical guidelines considerably reduces the element of judicial discretion in
sentencing and they have thus not been popular with judges (see Tonry, 1987, 1992
and 1993; Freed, 1992; De Benedictis, 1993).
These U.S. guidelines were all formulated by a sentencing commission or com-
mittee appointed by the government for the purpose. Sentencing commissions have
also been formed to provide advice for governments, e.g. in Canada, the state of
Victoria in Australia and a number of US states (Ashworth. 1992b). These have
had a very mixed reception from governments. The Canadian guidelines have not
been implemented while the Victorian Sentencing Act was passed in 1991.
3. Computer Support to Aid Sentencing Decision Making

Computer systems have been used both to support sentencing reform, and as a
reform in themselves. However, in developing such systems we need to pay close
attention to conceptualising the decision process of that domain. Oskamp, Tragter,
and Groendijk (1995) suggest that in AI and law research, "[t]he impression is
sometimes given that the (legal) assignment to be carried out by the system is of
secondary interest". (1995: 213) Instead, they argue, research needs to focus more
explicitly upon the appropriateness of different strategies of system modelling
for different domains and for different purposes (Oskamp, Tragter, & Groendijk,
1995). By implication, closer examination of the appropriateness of different strate-
gies necessitates the need to pay closer attention to the nature of that domain. In the
following section, I review briefly the attempts which have been made to develop
decision support systems for judicial sentencing focusing on the model of judicial
discretion upon which these systems have been based.
3.1. THINKING ABOUT MODELLING JUDICIAL DISCRETION

Bench-Capon (1994) and Zeleznikow and Hunter (1994) have argued in favour of
the development of computerised representation of legal rules to assist decision-
making. Zeleznikow and Hunter (1994) take the reader through a brief tour of
legal theory in relation to building intelligent information systems. They present a
debate between 'Legal Positivists' who argue that law can be represented by" ... a
settled body of rules ... " (p. 63) and 'Legal Realists' who take " ... a more extreme
approach and reject the categorisation of law as fundamentally about rules" (p.
53). Having acknowledged the centrality of the debate they introduce, the authors
conclude that " ... the majority of cases should be decided on something approach-
ing a positivist approach". However, this conclusion seems to be more as a result
of a pragmatic preference than of a theoretical argument, (other than to say that
strong legal realism is " ... an unlikely conclusion and one which many would not
[ lOl 1
206 CYRUSTATA
accept"). Why is there a temptation to prefer a positivist approach over other (such
as 'legal realism', or, sociolegal) perspectives? Alldridge (1997) suggests that in
thinking about the relationship between computers and law there is a tendency
" ... towards a highly formalistic positivism". He raises the question as to whether
this form of " 'legal reductionism' " (involving "if p then q") is, in one sense,
highly practical:
[T]he advantage in computers and law is that as soon as the 'legal' operation
is characterised as an if p then q operation, then it becomes easy to perform
it with a computer .... If you are going to use computers in legal contexts,
the pressure which computers generate is towards a rule-bound framework,
because it is what they do best .... The link is a contingent one - there is
no logical one, but a practical one. (Alldridge, 1997: 3-4, original emphasis
retained)
Alldridge argues that the most significant changes in law over recent years have
been brought about by computers. He argues that sociolegal study has not en-
gaged sufficiently with these developments and that they demand its consideration.
Although not dealing directly with decision support systems, Alldridge implies
that the challenge for the sociolegal tradition is to research and develop systems
which avoid the seduction offered by the practical convenience of positivist 'legal
reductionism' .
I would suggest that this positivistic, legal reductionist approach to decision-
support systems is founded on a restrictive view of the sentencing process as one
which is fundamentally prescribed by formal rules. There sometimes appears to be
a presumption in this positivist rule-based approach that 'rules' are (or should be)
formal legal rules, rather than behavioural rules (Hawkins, 1992). This presumption
ignores the inter-connection between social and legal processes during the criminal
process which set the agenda for the formal sentencing decision. Reliance on such
an approach may help to explain the difficulties which confront the design and
application of current rule-based systems. (See for example, difficulties reported
by Hassett (1993) in producing a 'Bail Advisor'). This is not to say that decision
support systems based on a positivist legal reductionist model do not have a role to
play, but simply that, as Oskamp, Tragter, and Groendijk (1995) imply, more criti-
cal attention needs to be paid to the application of such approaches to discretionary
judicial decision process. I would suggest that alternative approaches merit serious
consideration when we think about developing judicial decision support systems
even if the implementation of such approaches may still be relatively uncharted.
However, attempts have been made to move away from a simple positivist 'if p
then q' image of judicial sentencing. A hybrid approach involving both rule based
and case based systems has been developed by Bain (1989). The programme begins
with an empty case-library and a handful of heuristics for deciding sentences when
no cases can be applied to a new situation. After only a few cases, however, it
begins to retrieve 'remindings' of its own cases from memory and to modify the
strategies associated with those cases to form new sentences.
[102 ]
The idea of a case based reasoning system unsupported by heuristics has also
been used as a basis for modelling the sentencing process. Murbach & Nonn (1991)
report progress on a project to develop a sentencing support system for fraud cases
in Canada. Their system provides information about penalties but also includes in-
formation on case factors not included in the categories of offence used in the penal
code but agreed by judges to be relevant to sentencing. There is thus an attempt to
include information which reflects judicial perceptions of seriousness in order to
make the system more sensitive and thus more useful to sentencers. Computer tech-
nology has been used to assist these reforms and to encourage greater consistency
in sentencing. ASSYST (Applied Sentencing systems) has been developed by the
US Federal Court system so that criminal justice personnel could easily compute,
record, archive and examine the implications of the US Sentencing Commission
Guidelines (Simon & Gaes, 1989; Simon, Gaes, & Rhodes, 1991).
Schild (1995) reports work in progress to develop a case-based advisory sys-
tem for sentencing. The domain knowledge was elicited from a senior judge, and
the system uses "hierarchical discrimination trees" in order to retrieve relevant
information. Schild notes that:
[i]t is obvious that the area of sentencing is associated with an enormous
amount of both common-sense and domain knowledge. However, a model
which would include this knowledge was considered impracticable and so it
was therefore decided to use only the domain knowledge without any addi-
tional common-sense knowledge .. (Schild, 1995: 232)
A part-simple retrieval system, part-expert system approach to sentencing support
has been reported by Bainbridge (1991). The system focuses on sentencing prac-
tice in magistrate courts in England and Wales and contains components covering
sentencing law and penalty statistics. The sentencing law component is intended to
assist the magistrate by checking that the chosen sentence complies with relevant
sentencing law. This part of the system is arguably more like an expert-system than
a simple retrieval system, although it only answers the question, 'Is this sentence
legally competent?' rather than, 'What is the appropriate sentence for this case?'.
When a judge has selected a legally competent sentence, it is then possible to
consult the penalty information section of the system. This shows the distribution
of penalties for the offence in the form of probability calculations. Information is
only available for two statutory offences of theft and burglary and for only 600
cases from four magistrate courts.
Database technology has been used in a number of large scale information
systems. Such systems, generally referred to as 'Sentencing Information System',
have been implemented in a variety of jurisdictions and used in practical, day-
to-day basis to support the sentencing process. What is perhaps striking about
such systems is that appear, at least at first blush, to be simple databases. They
are not intended to carry out any retrieval process which is not transparent to the
judge; all the machine is doing is counting cases. However, as every statistician
knows: anyone can count but does s/he know what and how s/he is counting? First,
[ 103]
208 CYRUSTATA
however, it is useful to examine attempts to build, implement and institutionalise

such information systems.
3.2. SENTENCING INFORMATION SYSTEMS

A Sentencing Information System provides users with information about the range
of penalties which have been passed by the court for similar cases in the past. The
system allows the judge to enter certain information about the case which he is
considering and the range and quantum of penalties passed by the court for similar
cases is displayed.
Formally, a Sentencing Information System is descriptive rather than prescrip-
tive. That is, it contains no guidance as to how a sentencer might use this informa-
tion to help in making the sentencing decision in a particular case. A Sentencing
Information System (SIS) can display the range of sentences for the particular com-
bination of offence and offender characteristics selected. The sentencer is given
no instruction as to what extent and in what direction the appropriate sentence
for the case at hand should vary from the average. This decision is a matter for
the discretionary judgement of the sentencer. However, the frequency distribution
indicates the highest and lowest sentences previously passed for the type of case
at hand. In a well trodden area it might be assumed that a sentencer would have to
have good reasons for straying outside the upper and lower limits.
Jurisdictions in Canada and Australia have experimented with Sentencing In-
formation Systems. These are described briefly below.
3.2.1. Doob and Park System

One of the earliest systems was designed by Doob and Park in Canada and it op-
erated for six years in four provinces (British Columbia, Manitoba, Saskatchewan,
Newfoundland) (Doob & Park, 1987). By 1990 only the Saskatchewan system was
still in operation, in the latter stages, using only Court of Appeal information. Doob
has explained why he thinks judges did not make sufficient use of the system. There
are two main reasons. First, judges in Canada had little interest in information
about current court practice. They are not accustomed to using information in this
numerical form nor does their legal tradition give any weight to current sentencing
practice. Second, such authority as exists in sentencing comes from the Court
of Appeal. The Sentencing Information System carried no institutional authority
(Doob, 1990).
3.2.2. The British Columbia System

This system operated in British Columbia from 1987 to 1992 (Hogarth, 1988) but
is no longer operating. A private communication with the IT director suggests
there were two main reasons. First, there was insufficient judicial consultation and
involvement, partiCUlarly in the early stages of the project. Judges therefore felt that
[ 104]
the information provided by the system was not helpful to them. Second, the costs
of the system, although not revealed in detail, were very high. Schild (1995) has
suggested that Hogarth's system may be subject to criticism because " ... the statis-
tical knowledge embodied ... is based on a very small number of characteristics.
This does not suffice to express the actual complexity of the sentencing process".
He also criticises the structure of system: it quickly runs out of cases since there
are no hierarchies between 'variables'.
3.2.3. The Australian New South Wales Sentencing Information System

The New South Wales (NSW) SIS is now a discrete component of a larger ju-
dicial information system known as the 'Judicial Information Research System'
(JIRS) which also includes components on industrial, land and environmental law.
Nonetheless, it is probably fair to say that the SIS remains the most consulted
component. 2 The New South Wales SIS is connected to all NSW Judicial Officers
and contains a Court of Criminal Appeal Judgements Component containing over
3,000 full text judgements mainly dating from 1990. Cases can be retrieved by case
name or by a word search. Allied to this is a Court of Criminal Appeal summaries
component retrieving summaries prepared by staff of the Commission. A hypertext
link allows the user to view 'sentencing principles' database which is also prepared
by the staff of the Commission. The Sentencing Statistics Component is split into
Local Court data (with around 170,000 cases over the past two years, and Higher
Courts' (District and Supreme Court) data with around 20,000 cases over the last
six years. Data is drawn from the NSW Bureau of Crime Statistics and Research.
Other NSW SIS components include a Sentencing Law; a Facilities Component
(concerning the availability of various services and disposals), and a Sentencing
Date Calculator (used to calculate the exact date for the commencement and ending
of minimum and additional custodial terms) (Potas, 1997; Potas et aI., 1998).
The progressive expansion and development of the NSW SIS could be seen as
a sign that judicial officers have found (or discovered) a need for easy access to
systematic information about 'normal' sentencing practice and that the system is
both supported and regularly consulted by judges. There has, however, been no
programme of systematic evaluation of the extent of and nature of usage of the
system. At first blush, this seems rather surprising given the resources that the
system requires for development, maintenance, training etc. However, it could be
argued by the Commission that there are good reasons for delaying evaluation of
the nature and extent of consultation. 3 The NSW Judicial Commission has recently
re-engineered the system and, the argument might go, there would need to be a
2 Reports from the NSW Judicial Commission.

3 The Judicial Commission does however maintain evaluation material informally including for
example the number of 'log-ins' to the system and it is possible to see where and for how long the
system was used by any particular judge. (One judge has refused to use the system because of this
ability to track usage).
[ 105 ]
210 CYRUSTATA
period of around two years before a systematic programme of evaluation became

meaningful.
3.2.4. The Scottish Sentencing Information System

The initiative for this project came from the Lord Justice-Clerk, the second most
senior judge in Scotland, who had seen the New South Wales system demonstrated
in Canada at a Conference of the Commonwealth of Learning and decided that it
might be useful to have a similar system for the High Court of Justiciary. With the
support of the Lord Justice-General, The Lord Justice Clerk approached the Law
School at the University of Strathc1yde where there was existing relevant expertise.
The Scottish Office provided funding for a feasibility study which ran from June
1993 to January 1995 which produced a prototype. 4 The operation and description
of the taxonomic issues is documented elsewhere (e.g., Hutton, Tata, & Wilson
1995; Tata, Hutton, Wilson, Paterson & Hughson, 1998).
The Scottish Courts Administration is currently funding (from November 1995)
the first phase in the implementation of the prototype. This involves the delivery
of the system to up to fifteen judges over the course of one year and on-going
study of how the system functions when operated within the day-to-day judicial
and sentencing environment; as well as its utility so as to identify weaknesses and
areas needing further development (see Tata, 1998a). This first phase of implemen-
tation has also begun conceptual work on an Appeal Court Judgements Database
to be subsequently incorporated into the SIS (Tata, Hutton, Wilson, Paterson, &
Hughson, 1998). Perhaps a distinctive feature of the Scottish SIS is that the data
which it uses is not drawn from official sources but is collected specifically for
the system. This allows a relatively high degree of flexibility and specificity in
the representation of data. However, not only must decision support systems deal
with the source, quality, and structure of data, but perhaps equally important is
institutional acceptance and support to which I will not tum.
3.2.5. The Institutionalisation of Information Systems

In its White Paper (Scottish Office Home Department, 1994), the government wel-
comed the Sentencing Information System initiative and indicated that if it was
workable and successful, similar systems might also be considered for the lower
courts. It would therefore appear that there is some government support for sen-
tencing information systems. While the future of the Scottish SIS seems to enjoy
generally strong judicial support there are perhaps two broad issues which will
require further attention. First, unlike the New South Wales SIS, an operational
Scottish SIS may well need to enjoy direct institutional protection. There does not
yet appear to be any long-term plan for the system to be run by an official body
4 This was developed in close consultation with the High Court jUdiciary. Regular and fre-
quent meetings were with a 'judicial sub-group' and successive early prototype versions were also
presented to 'judicial seminars'. See Hutton, Paterson, Tata, and Wilson (1996) for fuller details.
[106 ]
directly responsive to judicial needs. In New South Wales, on the other hand, the
SIS is administered by the Judicial Commission of New South Wales. In Scotland,
however, the possible absence of an official body with a direct interest in defending
the system from criticism and resource pressure may leave it vulnerable. Secondly,
given this possible future vulnerability to public and resource pressures, it may
be difficult to lobby for funding to improve and expand the system in response to
judicial requests for more or different information. As with any computer system,
the SIS will require to evolve and improve over time. The apparent ability of the
New South Wales system to do this may help to explain its longevity and apparent
success compared with its North American forebears.
Closely inter-related with the question of institutionalisation are the questions
of the 'impact' and institutional authority of systems.
3.2.6. Impact and Institutional Authority

One of the most intriguing questions which will need to be explored will be the
extent to which the introduction of information technology will change the nature
of sentencing practice. Although there may be no formal requirement for judges to
consult the system and take note of its information, there are reasons to suppose
that judges may feel informally encouraged to use the system and indeed that it
may significantly impact on judicial sentencing behaviour. Indeed, the very process
of building such systems in close consultation with judges means that systematic
discussion of issues and practice is disseminated within the judicial community.
(For further explanation of this point see Hutton, Tata, & Wilson, 1995).
Some commentators have suggested that, given the relative dearth of systematic
information about sentencing practice ')udges ought not only to be provided with,
but would positively delight in, access to detailed information [about sentencing
practice]" (sic, Zdenkowski, 1986: 232; see also Ashworth, 1997; Hedderman &
Gelsthorpe, 1997). However, Doob's reflections on judicial use ofthe system which
he developed in Canada are sobering. He has described the slow "closing out"
of that project and questioned the assumption that (Canadian) judges, far from
'delighting' in knowledge about current practice, actually' ... want to have easily
accessible to them knowledge of current sentences being handed down in com-
parable cases ... ' (Doob, 1990: 2). Indeed the claim that judges report that they
are more concerned with 'internal' or personal consistency than with inter-judge
consistency has been documented by other studies (Tarling, 1979; Hutton & Tata,
1995). The reason for this, Doob argues, is simply that judges operate within an
environment which does not reward attention to "current practice". Essentially,
Doob has argued that from his experience, and also (he implies) from the similar
fate met by Hogarth's system, judges do not perceive there to be a need for this
kind of information about 'normal practice'. Since judges cannot be coerced into
paying attention to such information systems, then judges do not perceive a 'need'
for such information. The crucial determinant of authority, Doob suggests, may lie
in the legitimation of such systems by the Appeal Court:
[107 ]
212 CYRUSTATA
Our experience in Saskatchewan, where the project seems to be most success-

ful, suggests that it is important to have the system 'legitimized' by a part of
the system that has the authority to do this. I am reasonably sure that it is very
important that the prime authority for our involvement in Saskatchewan is the
Court of Appeal. (Doob, 1990: 12)
To what extent has this institutional involvement been possible in Scotland and
NSW? In Scotland the initiative for the SIS came from the Lord Justice-Clerk, the
second most senior judge in Scotland, with the support of the Lord Justice-General
and a number of other senior (Appeal) judges. However, while it has enjoyed this
informal senior judicial support it is too early to say whether it will enjoy the
more formal institutional support of the Appeal Court. The system has not yet
been referred to by the Court of Appeal and its current status is as an informal,
private judicial aid (Tata et al., 1998). While the initiative for the system in NSW
appears to have been rather different (the Judicial Commission and its system were
initially imposed on the judges (see for example, Basten, 1980; Basten, 1995)), it
is now beginning to be possible to discern the formal institutional attitude of the
Appeal Court towards the NSW SIS. Wetherburn & Lind (1996) argue that given
the traditional reliance on appellate review in sentencing, " ... the response of the
NSW Court of Criminal Appeal is likely to prove to be pivotal in determining
its usage". (Weatherburn and Lind 1996: 162) They remark that, "[a]t this stage
the court's attitude to the SIS could best be described as lukewarm" (Weatherburn
& Lind, 1996: 162), although more recent examination perhaps reveals that the
picture may be more mixed than this and may be becoming 'warmer' (Tata, 1998b).
However, looking at the introduction of information systems with a rather more
critical eye, a "lukewarm" judicial attitude in New South Wales can at least partly
be explained by a limitation in the ability to retrieve what are seen by judges to be
genuinely 'similar' cases. Arguably this weakness is, at least in part, derived from
the use of official (typically police) data sources (Tata, 1997: 397-401). While the
institutional support (particularly from the Court of Criminal Appeal) may be one
key factor in ensuring judicial acceptance and use, another crucial factor relates to
how judicial intelligence about sentencing is modelled and represented, an issue to
which I shall now turn.
4. How can the Representations of Knowledge Enhance the Utility of

Decision-Support Systems for Sentencers?
Aside from questions of political support and the relationship between information
systems and judicial culture, a key question must concern the conceptual char-
acter of a system intended to support decision-making so strongly characterised
by formal discretion. Although in one sense every case is unique (because each
individual, social situation and context are unique), it is possible, at least in prin-
ciple, to compare cases and therefore to say that some cases are 'similar' (Hood,
1962). The critical question, then, must be: 'how is 'similarity to be understood?'
[ 108]
Traditional academic understanding has largely been informed by 'legal analysis'.

Below I argue that to-date attempts to construct systems of computer support for
sentencing decision-making have tended to be rooted in a perspective which is
primarily legalistic in outlook and analytical in its thinking. I would argue that
systems based on this paradigm offer at best only a restricted understanding of
judicial discretionary decision-making and therefore computerised representations
modelled on such a perspective will be of only limited intelligibility to judges.
4.1. THE WEAKNESS OF THE LEGAL-ANALYTICAL PARADIGM
What are the characteristics of the 'Legal-Analytical' paradigm in its attempts to

represent 'similarity' and what limitations does it have? Let us consider 'legal
reductionism' and 'analysis' in tum.
4.1.1. Legal Reductionism

In developing decision-support systems for judicial sentencing, the starting point
for representations of 'similarity' has tended to be a legalistic one (Bainbridge,
1991; Doob, 1990; Gruner, 1991; Henham, 1997; Hogarth, 1988; Potas, 1997).
Systems search for 'like' cases or representations of 'like' cases primarily in terms
of official legal convictions. However, in designing the Scottish SIS the appro-
priateness of such a starting point became increasingly unclear. From work with
judges (Hutton, Paterson, Tata, & Wilson, 1996) and previous research with sen-
tencers (Ashworth et aI., 1984; Hogarth, 1971; Hutton & Tata, 1995; Wilkins et aI.,
1978) there was strong reason to doubt the view that judges think primarily in terms
of classifications represented by official headline categories of the official criminal
law 'headline' offence. Previous research (Fitzmaurice & Pease, 1986; Ashworth et
aI., 1984; Hood, 1962; Hood, 1992; Hutton & Tata, 1995), suggests that sentenc-
ing is determined by far more than the legal definition of the offence(s). Judges
consider the circumstances surrounding the commission of offences as important.
However, can the famous elasticity of common law offences or the famed precision
of 'statutory offences mean that they would be sufficient building blocks for a
decision support system? I argue below that neither would tend to be appropriate.
As in other countries where sentencing is conducted in a predominantly com-
mon law jurisdiction, in Scotland the strict common law headline offence category
often provides little information as to the seriousness of the offence from the
perspective of sentencing. A sentencer must not only consider the common law
conviction, but also the circumstances of the events surrounding the commission
ofthe offence. The 'headline' offence(s) (for example, 'Robbery', 'Rape', 'House-
breaking' 'Theft' etc.), with which an offender is charged and convicted may be
of limited relevance in deciding sentence. Very frequently, the common law head-
line conviction does not provide a sentencer with sufficient information about the
circumstances surrounding the commission of the offence.
[109 ]
214 CYRUSTATA
If common law offence classifications provide insufficient information about

the circumstances of an offence and its attendant seriousness, it might be thought
that statutory based offences provide sufficient precision for the consideration of
sentence. After all, could the organisation of information not simply replicate all
offences as they appear in Acts of Parliament? Aside from the practical considera-
tions of operating such a system, there is a conceptual difficulty. While the creation
of offences on the statute books are used to charge persons, they may often be of
limited assistance when judges consider sentence. It is not clear how to organise
statutory offences. One possibility would be simply to reproduce every section
or even every subsection of every Act of Parliament relating to sentencing. How-
ever, this typically means that 'similarity' is determined less from the perspective
of judicial sentencing and rather more from the perspective of the parliamentary
draftsman (Hutton, Paterson, Tata, & Wilson, 1996).
An alternative course of action would be to arrange offence information in a
way which is relevant from the perspective of sentencing. This might consequently
mean that the arrangements of offences would not necessarily reflect the divisions
within the criminal law. Thus the assumption that a decision support system's
taxonomy should begin with the official criminal law headline offence category
is very doubtful. Rather, the single most important criterion which judges consider
is not the official criminal offence category, but its relative 'seriousness' (Wilkins
et aI., 1978; Fitzmaurice & Pease, 1986; Hutton & Tata, 1995). Although decision-
support systems for sentencing have recognised the need to incorporate information
relating to offence seriousness, the organisation of that information has still tended
to privilege a legalistic starting point.
4.1.2. Analysis
While representations of similarity have tended to assume a formal legalistic start-
ing point, conventionally there has been a recognition that there needs to be some
account taken of factors which 'aggravate' and 'mitigate' the seriousness of the
headline conviction. Typically, having selected the 'appropriate' criminal law clas-
sification the judge is then invited to add in standard aggravating or mitigating
factors (Chan, 1991; Hogarth, 1988; Murbach & Nonn, 1991). What is striking
about this approach is the additive and analytical nature of the representation of
similarity. I would suggest that this analytical process of abstracting additive 'fac-
tors' from the whole case is an artificial representation of similarity. The abstraction
of independent factors which purport to describe seriousness denies the relational
meaning of information in a case. It assumes that a case can be meaningfully frag-
mented into discrete and abstract individual factors which can be analysed as if they
each have a power independent of their own over the decision process. However,
'aggravating and mitigating factors' only make practical sense to the sentencer in
relation to each other and to the whole case.
As an illustration, the reader is invited to consider one issue which has received
surprisingly little attention: cases where an offender is convicted of more than one
[110 ]
charge. The legal-analytical approach would try to record each conviction sepa-
rately, or, just one 'main' conviction. However, this does not appear to be they way
that sentencers tend to think about cases (Hutton, Paterson, Tata, & Wilson, 1996).
In considering sentence, judges do not seem to think about each conviction in a
multi-charge indictment in isolation from each other and from the circumstances
of the commission of the offences. Rather, they tend to view the case as a whole
incident or narrative of events (AI schuler, 1991; Parton, Hansel, & Stratton, 1991).
Therefore, in attempting to reveal the process of the decision-making of sentencers,
empirical research has used sentencing vignettes (Ashworth et al., 1984; Corkery,
1992) rather than necessarily relying on criminal law classifications which are
necessarily limited in the relevant information which they can offer the sentencer.
While the legal-analytical approach to representing 'similarity' of cases for the
purposes of sentencing may be limited, there is a possible alternative perspective
which I suggest below.
4.2. A SCHEMATIC AND HOLISTIC REPRESENTATION OF 'SIMILARITY'

Judges frequently tum to the cumulative and evolved wisdom of judicial experience
represented by precedent for their decisions and also place particular emphasis on
their personal experience. This 'experience' helps judges to interpret and schema-
tise the mass of information with which they are presented. In developing this
argument, let us consider research into discretionary legal decision-making more
generally.
It has been established by cognitive psychology that humans are limited proces-
sors of information. From his field experimental research into the psychology of
prosecutorial decision-making in criminal sentencing, van Duyne (1987) found
that despite the considerable discretion which it afforded, prosecutorial decision-
making could be characterised as " ... one dimensional: the Prosecutor selected
out of the total information on the case only those aspects which were consistent
with a particular 'dimension' (e.g., 'professional', 'social misfit' or rehabilita-
tion) and fitted these into simple conceptual schema". (at 147) Giller and Morris
(1981) found that social workers use 'operational philosophies' (the means by
which professional ideologies are mediated through the demands of practice):
Having located the moral character of the case, the social workers were able
to respond with a repertoire of provisions which routinely met the case as
portrayed. In this way, social work with offenders became ordered and rational
and a work priority was established. Decisions were not 'made'; they emerged
as natural logical, even inevitable . .. (at 79-80)
If discretionary decision-makers are quickly able to characterise a case before them

by making use of simple conceptual schema, why is it that discretionary decision-
makers so often report experience of 'difficulty' with cases? Van Duyne (1987)
concludes that the results from his study suggest that "[iJf sentencing is difficult,
[ 111 J
216 CYRUSTATA
it is because of its uncertainty, not because of its complexity .... In problem

solving, (it is) necessary to restrict the information to one manageable dimension
in order to avoid overload and uncertainty ... " (at 151-152). Lawrence (1988)
reports similar findings about the inferences magistrates made about available data.
"Expert-novice differences were in the inferences they made ... [E]xperts built up
a picture of [the offender] or a patterned expectation ... " (at 241). In a similar vein
Emerson (1995) argues that" 'seriousness' becomes routinized, institutionalized,
built into the typification, as it were, rather than standing as an experiential feature
... (at 167).
The model that decisions are in practice taken through a legal-analytical process
characterised by a sequential, linear process comprised of formal and deductive
reasoning is doubtful. Rather like other problem-solvers, lawyers and judges make
use of experience to help them to schematise new cases. For example, Crom-
bag, Wijkerslooth and van Serooskerken (1975) asked experienced legal problem-
solvers to think aloud while solving a concrete problem. They found that,
[t]he most striking result was that what was said while thinking aloud created
a rather chaotic and unsystematic impression. Often a person seemed to have
a solution, although a provisional one, at an early stage for which he subse-
quently tried to find supporting arguments. Moreover, during the reasoning
process, the subject did not seem to complete one part after another, but rather
to jump wildly back and forth (at 169)
There has tended to be an assumption fostered both by the opinions delivered by

judges in court (and sometimes reported by law journals) and also taken up by
legal writers and researchers (e.g., Fox, 1994; Hogarth, 1971) that cases are most
logically understood by breaking them down into two main components. As we
saw above, the seriousness of the offence is often said to be broken down into
legalistic classifications with aggravating and mitigating factors (e.g., Henham,
1997; Lovegrove, 1989; Moxon, 1988). Similarly, a conceptual distinction is made
between the offence and 'the offender'. So for example, it appears to have become
received wisdom that it is normal practice for sentencers to begin by considering
the criminal law conviction, then balance this according to the weight of 'aggra-
vating and mitigating factors' and then tum to consider the previous convictions of
the offender and finally his or her personal circumstances (e.g., Fox, 1994; Boyle &
Allen, 1990). Rather, sentencing is better understood as an intuitive and relatively
impressionistic process, but is not necessarily any less reliable for that. Indeed, Chi
(1988) suggests that one of the differences between novices and expert decision-
makers is the ability of the expert to arrive at a decision quickly but yet consistently
and defensibly. Novices on the other hand tend to try to come to a decision ac-
cording to a rather formalistic, linear, sequential and analytical structure. Expert
decision-makers are in fact more reliable and consistent in their decisions since
they are able to process information and make judgements on the basis of informal
intuitive structures of 'knowledge' and associations.
[ 112]
Therefore, a more schematic and holistic approach would attempt to classify

different narrative types of criminal conduct exclusively from the perspective of
sentencing. In the shape of this more schematic-holistic approach there may be the
ability for judicial decision support systems to capture the practical nature of dis-
cretionary decision-making rather than positing it as a sequential, legal-analytical
set of discrete stages. The idea behind a schematic-holistic approach is to attempt
to derive classifications from the mental schema around which judges, like all
decision-makers, tend to operate. Thus, this approach would aggregate or split legal
constructs, or, indeed 'invent' terms to sum-up informal judicial intelligence.
4.3. IS A SCHEMATIC-HOLISTIC APPROACH SIMPLISTIC?

It might be objected that the notion of a computer system which retrieves cases on
the basis of a very small number of selections is too crude. Since it does not rely
on official criminal law categories as a necessary starting point, it might be argued
that such an approach overly-simplifies the intricate knowledge represented in each
case by the criminal law. The attempt to produce such a typological representation
of information about offences may be thought to be overly simple since the very
idea of 'types' of cases denies the delicate and subtle complexities of the 'knowl-
edge' represented by the criminal law. However, to make such an argument is to
lose sight of the fact that this officially defined 'knowledge', or 'reality' is itself
constructed for a particular purpose: securing convictions and acquittals. All crim-
inal convictions are themselves typologies (Sudnow, 1964). Categories of criminal
convictions do not, and necessarily cannot, reflect the full precise occurrences of
'what happened' during a criminal incident or series of related incidents (Bennett
& Feldman, 1981). Rather, the criminal conviction for which a person may be
sentenced is necessarily the result of a simplified and normalised representation of
events. (Shapland, 1987; Curran & Chambers, 1982; Giller & Morris, 1981).
Both Shapland (1981) and Ashworth et al. (1984) suggest that what legal rhetoric
refers to as the 'facts of the case' should not properly be regarded as 'facts' but as
" ... constructions of the circumstances forming the offence, stemming from the
constructions made by the police (themselves very much negotiable and influenced
by the individual views of detectives and the wishes of victims and witnesses)
(Ericson, 1981; Ericson, 1982)". (Shapland, 1987) Thus the criminal process brings
to the sentencer a typified stream of cases which intuitively suggest typical and
schematic meaning about the nature and seriousness of the case. These provide the
sentencer with cues about how the criminal process has constructed the serious-
ness of the case and its expectations of the sentencing outcome. From this vantage
point of the criminal process, it becomes clear that the formal judicial sentencing
decision is only one decision in the overall processing of a case for sentence. That
decision is based on typical constructions of cases by their flow through the whole
criminal process. Therefore, the attempt to analytically unscramble the judicial
sentencing decision into individual factors so as to try to represent sentencing as a
[ 113]
218 CYRUSTATA
formally sequential and deductive process is naturally likely to encounter judicial

resistance.
What then of the attempt to model judicial thinking about criminal history ?
4.4. JUDICIAL INTELLIGENCE AND CRIMINAL HISTORY

How are understandings of similarity concerning criminal history represented by
systems of decision support? The NSW system posits four groups of criminal his-
tories (derived from official data): no prior convictions; prior convictions but not
of the same offence type; prior convictions of the same type; prior convictions
with imprisonment. 5 Currently, the Scottish SIS records and retrieves informa-
tion according to three dimensions: 'Jurisdiction' (whether previous convictions
were under solemn or summary jurisdiction6 ); 'nature' (whether they were broadly
'analogous' or 'non-analogous,7); sentence type (whether they received a custodial
or non-custodial sentence). To each of these dimensions are recorded the options
of 'yes', 'no' or 'unknown'. In retrieval judges can also choose the default option
'ignore' which retrieves 'yes', 'no', and 'unknown'. These 'factors' therefore com-
bine to form a matrix of choices for the recording and retrieval of information about
criminal history (Tata et aI., 1998). How adequate are such taxonomies, (which rely
on a limited number of 'factors'), as a basis for representing judicial intelligence
on the seriousness of previous criminal history?
It may initially seem that the dimensions used in the Scottish system cannot
record enough qualitative information about the nature of an offender's previous
convictions because it only provides associative measure of seriousness. However,
it should be recalled that associative information is all that is available to the judge.
In other words, judges normally have to make a judgement about prior criminal
record on the basis of association of terms such as 'solemn'; and the bald criminal
law conviction. Schedules of previous convictions are the major source of informa-
tion about prior criminal history and the information they list can only provide an
association with the prior case before the previous judge rather than a full replay
of the offender's life.
However, judges also find that this matrix of three dimensions provides them
with inadequate information about the seriousness of the record. They are also
interested in recording and retrieving information about a multitude of dimensions
such as the length of previous custodial histories; the number of convictions of
particular types of offences; the chronology of offending and custodial institution-
alisation and so on, and the inter-relationship between these dimensions. However,
as it already stands the specification of the previous convictions characteristics dra-
5 This four-fold taxonomy applies only in the 'higher courts' module. In the local courts module
there is a simple distinction between 'priors' and 'no priors'.
6 'Solemn' jurisdiction tend to deal with more serious cases and necessitate a jury trial, whereas
summary cases do not involve a jury and tend to deal with relatively less serious cases.
7 The definitions of what is 'analogous' and 'non-analogous' were decided by the judges.
[ 114]
matically reduces the number of sentences which are retrieved by the SIS. Judges
have reported that selecting three or four combinations of previous convictions will
often reduce the number of cases retrieved to zero and that this can become frus-
trating. There are potentially, 6,561 possible combinations of previous convictions
alone (a figure which is equivalent to the number of sentences passed over seven
years which is held by the system). No wonder then that this is the gate through
which judges find that they frequently lose potentially similar cases!
So on the one hand we have the justifiable complaint that such a matrix is
overly-specific and complex; and yet on the other hand, there is the equally justifi-
able complaint that there is not enough information about seriousness. Is this just
another instance of judges' unquenchable thirst for the availability of information
which then leads to a frustrating hunt for similar cases (Hutton, Tata, & Wilson,
1995)? In thinking about possible resolutions of this apparent dilemma, there are
three options which might be explored. However, as I shall argue none of these
provides a satisfactory model of judicial 'knowledge' about criminal history. An
alternative perspective will then be offered.
Option A: Groupings arranged according to broad seriousness
A simpler classification than the existing matrix may be presented as say three
broad groupings of previous convictions according to their relative seriousness.
This might work according to groups of cases according to broad bands which
relate to their relative seriousness. For example, previous sentences where the of-
fender had a record of solemn analogous custodial convictions might be grouped
under the 'very serious previous convictions', while sentences where the offender
had no solemn, no analogous of custodial sentences might be grouped under the
'less serious previous convictions' classification. Obviously, these are two rela-
tively straightforward examples of cases at the 'extreme' ends of the spectrum and
it would have to be for judges to decide a workable set of groupings for the purposes
of the system.
The benefit of this kind of grouping is that it would allow judges to retrieve cases
with previous criminal convictions of very broadly similar seriousness without the
frustration of loosing large numbers of cases.
Option B: The simple reduction of information available
A second option to make the existing matrix less complex lies in the simple re-
duction of 'factors' present in the matrix. Thus for example, the ability to include
only those cases where the offender has prior convictions under summary jurisdic-
tion, or where the offender received a non-custodial sentence could be completely
removed. The obvious problem with this is that the simple reduction of the num-
ber of factors does nothing to address the complaint that matrix does not provide
sufficient information to gain a sense of the seriousness of criminal histories. In-
[ 115 ]
220 CYRUSTATA
Table I. Number of cases retrieved for a sample using AND and OR
Previous conviction settings No. of cases
Summary analogous custodial AND Solemn analogous custodial 3

Summary analogous custodial OR Solemn analogous custodial 13
deed, it has not proven possible for judges to decide which of the three dimensions
(,jurisdiction', 'analogusness', 'custody-non custody') should be sacrificed.
Option C: Logical operations 'AND' and 'OR'

The Scottish SIS is based on the assumption that factors identified by the user
as important in a case are combined (i.e., 'ANDed') together. Experience suggests
that, in general, this is a valid assumption. An exception to this may be the incorpo-
ration of previous convictions in the case details. Combining a number of previous
conviction categories using an 'AND' operator will quickly limit the number of
cases returned to the user. Using an 'OR' operator to combine the same character-
istics causes no such restriction. For example, in the case of an assault involving
a firearm and a victim vulnerable by dint of his/her occupation, the numbers of
cases retrieved when previous conviction choices are 'AND'ed' and when they are
'OR' ed' are shown in Table 1.
While this may appear to provide the judge with greater flexibility in how s/he
constructs searches of the database, it is still felt by judges to go not much further
forward in the ability to capture the seriousness of criminal records. The problem
is felt by judges to lie in the insufficiency of available dimensions and crucially the
inter-relationship of those dimensions.
Option D: Statistical modelling of previous criminal history

It would also be possible to construct a model of sentencing behaviour in dealing
with criminal history on the basis of previous judicial practice (e.g., Lovegrove,
1989, 1997) through for example the use of multiple regression analysis and a
model of statistically weighting various criminal history 'factors'. 8 The difficulties
8 Using such an approach it might also be possible to employ the techniques of Case-Based Rea-
soning (e.g., the 'Judge' System (Bain, 1986, 1989). As I argue below it is not the fundamental aims
of Case Based Reasoning or Neural Networks (Warner, 1990), or, indeed a combination of a Rule-
Based System and Neural Networks (Zelesnikow, Stranieri, & Gawler, 1996; Ze1eznikow & Stranieri,
1997); which are necessarily problematic, but I rather I am concerned to show that it is possible to
conceive of an alternative (and perhaps in this 'domain' more appropriate) way of modelling the
discretionary decision process rather than rather formalised and deductive analysis. (See also the
difficulties discussed by Edwards and Huntley (1992) in attempting to produce a rule-based reasoning
approach in the discretionary area of Family Law).
[116 ]
with this broad approach, as I argue below, are two-fold. First, it assumes that it is
possible to split up a judge's understanding of criminal history (in the context of
the whole case narrative) into independent factors or variables which are posited
as having a universal, discrete power over the sentencing decision. While it is of
course possible to demonstrate that there is a statistical association between the
presence of 'variables X and Y' and 'outcome Z', causality can only be inferred.
Thus, a second difficulty inherent in this approach is that it inevitably requires
arbitrary judgements to be made as to what constitutes the 'variables' (how one
'variable' can be distinguished from another etc.).
One promising approach to resolving the danger of fragmenting the meaning
criminal history could be to employ a Neural Network. This would enable the de-
cision support system to learn the weights of each of the relevant attributes of crim-
inal history in sentencing outcomes. In this way the neural network 'learns' the sta-
tistically expected combinations between different factors (Zeleznikow, Stranieri,
& Gawler, 1996; Zeleznikow & Stranieri, 1997). While the concept of neural net-
works is refreshing in the way in which it takes seriously the inter-relationship
between different characteristics and recognises that typical patterns may emerge,
in common with the Options A-C outlined above, it nonetheless begins from the
starting point that information about criminal history (indeed the whole case nar-
rative) can be dissected into discernable pieces of meaningful information. It must
begin by trying to fragment a case into identifiable abstract and discrete 'bits' of
criminal history as if they enjoyed meaning and power when abstracted from their
case context.
Below I suggest a possible alternative perspective from which we might think
about modelling previous criminal history in the context of the 'whole case' .
4.4.1. Modelling Qualitative Judicial 'Knowledge' about Criminal Histories

None of these options outlined above resolves the problem of the qualitative un-
derstanding which judges have of prior criminal records. Judges often report that
they do not necessarily analyse a schedule at-length but get 'a feel' of the kind of
offender with whom they are dealing by asking themselves simple questions such
as 'is this man salvable or not?' Experienced judges quickly process a great deal of
information not as isolated components or details but in relation to each other (e.g.,
Hedderman & Gelsthorpe, 1997: 55-57), so that they build up a meaningful picture
of the criminal experience of the offender: a 'feel' for the criminal life history of
the offender.
Thus it would be fruitful to explore with judges the possibility of identifying
'typical' criminal histories (e.g., see Parker, Sumner, & Jarvis, 1989; Allen, 1987);
for example, 'the repeat petty offender'; 'the one-off serious violent offender'; 'the
petty economic offender'; 'the large scale crook'; 'the young first time offender';
'the persistent child sex abuser'; 'the petty drug dealer'; 'the organised drug dealer'
and so on. Doubtless, such typical classifications would be different to, fuller and
richer than the examples given here, but it seems that judges do build up notions
[ 117]
222 CYRUSTATA
of typical kinds of criminal lives. It may be that these typical schematic classifica-
tions are consonant with the criminological literature relating to studies of criminal
histories (e.g., Barnett et aI., 1992; Greenberg, 1991). However, that is not really
the issue here. Rather, it is the attempt to try to represent judicial understandings of
types of criminal histories which capture the sense, (if not the futile quest to capture
the entire detail), of the information which routinely comes before sentencers.
It is suggested that the tabular information contained in the schedules of pre-
vious convictions should not be represented in isolation, but rather the schedule is
imbued by the sentencer with relational meaning by other information submitted
to the court such as Social Inquiry Reports, Pre-trial Reports, Psychiatric Reports.
All of these assist the sentencer in identifying the broad type of criminal history.
However, they do not just build up pictures of the criminal histories of different
types of offenders, but they also help to calibrate and explain the seriousness of the
offence at-hand. The pictures built up about criminal life stories shape and contex-
tualises the understanding of the offender's behaviour. Similarly, Lloyd-Bostock
(1988) has observed,
Prior record provides information on how set the offender is on a criminal

career. This may lead to a moral judgement that the individual is more (or less)
wicked, and hence more or less deserving of a more (or less) severe sentence.
It may also be seen as one dimension of the concept of responsibility. (at 73)
Prior criminal record tells the sentencer about the meaning of the offence at-hand
and motivation of the offender and his responsibility and culpability in the offence-
at-hand. If it is true that judges build up an intelligent and meaningful picture of
both 'the offence' and 'offender history' then representations of sentencing need
to try to reflect this. Thus, what is required is not just schematic-holistic represen-
tations of 'offence' and then 'offender', but rather an integrated understanding of
short typical narrative& of the whole case.
From this perspective decision support systems could record and retrieve cases
by attempting to capture qualitatively judicial 'knowledge' of a limited number
of different types of typical sentencing vignettes. The task for researchers and de-
signers of systems to support sentencing, and perhaps other areas of discretionary
decision-making, is to try to describe and represent this decision process which
I have characterised as one which constructs recognisable 'schema' in which the
context of the whole case is critical to its meaning (i.e., schematic-holistic).
However, can this be done? Two potential difficulties might be encountered.
First, a schematic-holistic approach is informed by a recognition of the way in
which cases are constructed and patterned in recognisable ways. However, is this
view of the criminal process (which sets the agenda for the sentencing decision)
one which is recognisable in a variety of jurisdictions? The second potential prob-
lem is that it might be felt that this perspective of the sentencing decision as a
process of patterning whole cases is too intangible to be modelled by computer
technology. Below I deal first with the question of jurisdictional specificity and
[118 ]
then with the question of whether in principle such an approach can be modelled
by computer.
5. Can it be Done? (Is a 'Schematic-Holistic Approach' Applicable and

Open to Modelling by Computer Technology)?
5.1. JURISDICTIONAL SPECIFICITY?
Let us first tum to the question of jurisdictional specificity: is this 'schematic-
holistic' perspective to modelling judicial intelligence applicable to continental,
civil law systems? It might be argued that although this process of the construc-
tion and re-construction of the typification of cases may be a reasonable portrayal
of Anglo-American adversariallegal systems rather than inquisitorial continental
European systems. However, in their comparison of the German and US systems
of criminal process Frase & Weigend (1995) conclude that "the two systems of
criminal justice appear to be converging towards a single model that incorporates
both adversarial and 'inquisitorial' elements". (Frase & Weigend: 318) Similarly,
the Dutch scheme is becoming increasingly 'adversarial' .(Jorg, Field, & Brants,
1995).
However, whether or not it is true that " ... Western systems of criminal Justice
are similar in a number of important respects and may be converging" (Frase &
Weigend, 1995: 359), my argument is that no system of criminal process can find
the 'objective truth' about a criminal incident, but must inevitably interpret and
re-interpret, actively historicise and re-historicise behaviour in a way which it can
comprehend. It can only perform this function according to a necessarily limited set
of typical schematic case constructions. These constructions necessarily undergo
a process of typification and standardisation through the criminal process which
in tum sets the sentencing agenda (Tata and Hutton, 1998). I have tried to show
that it is in the nature of (legal) discretion that the decision-maker needs to limit
and typify the process of trying to make sense of 'new' information so that it is as
similar as possible to a limited range of typical case scenarios. Rather than trying to
analyse and dissect each element of an individual case, ('offence' from 'offender';
the seriousness of each individual conviction; 'aggravating and mitigating factors'
etc.), an alternative approach would be to try to identify typical narrative case
constructions. Such an approach would accept these case constructions as whole
entities; rather than trying to break them down into discrete parts the cost of which
is the meaning of the whole case.
5.2. CAN SUCH A QUALITATIVE PERSPECTIVE BE MODELLED? (RE-THINKING

RULES AND DISCRETION)
There has been judicial resistance to the attempt to represent sentencing accord-
ing to some mathematical model associated with the addition and subtraction of
'independent' 'factors' (Lovegrove, 1989; Kapardis, 1987; Wilkins et aI., 1978).
[ 119]
224 CYRUSTATA
Although this might be explained by the invocation of the concept of 'judicial

independence' (e.g., Armytage, 1995), it can also be understood in terms of the
difficulty in trying to account logically and analytically for their decisions. Judges
stress their treatment of cases as a 'whole' and an intuitive 'feel' for individ-
ual cases. Does this mean that the use of a schematic-holistic perspective to try
to model decision support in sentencing is in fact too informal, amorphous and
intangible to be modelled by a computer?
In thinking about the possibility of modelling discretion for a decision support
system without reliance on reasoning justified in official sources, an (understand-
ably) pessimistic view appears to be taken by writing in the AI and law field.
Zeleznikow & Hunter (1994), for example, state that while a 'positivist' or legal
reductionist perspective enables the easy production of decision support systems, if
according to "legal realists ... each decision is made according to a series of factors
not articulated in the judgements and not amenable to modelling, then we can
never hope to use computers in legal reasoning". (Zeleznikow & Hunter, 1994: 63)
Similarly, in trying to answer the question of whether computers can understand
legal reasoning, Tito (1987) succinctly states that, "[c]omputers can only do what
humans program them to do". However, Tito presents two views of how lawyers
decide: "[Do lawyers] ... apply rules to decide when two situations are similar
within a legal context? Or, do lawyers just get a 'hunch' or have a flash of insight
that alerts them to similarities?" (original emphasis retained, Tito, 1987: 411) Tito
presents the latter view as "ruleless" and therefore not amenable to understanding
by a computer because it is not understandable to humans. However, I wish to
suggest that it is possible to argue that legal decision-making may be both based
on an 'instinctive hunch' and amenable to 'rules'.
Like other legal-analysts, Tito's conception of 'rules' seems to ignore the pos-
sibility of non-legal rules of behaviour (Hawkins, 1994). Likewise, Dworkin, for
example, neatly envisages discretion as "like the hole in a doughnut, does not exist
except as an area left open by a surrounding belt of restriction" (Dworkin, 1977:
31). However, one danger of this formulation is that it leads us to regard social be-
haviour as being regulated only by 'law' and that without law behaviour is (or must
inevitably be) unpatterned. However, as Lempert writes " ... discretion is not only a
property of legal rules it is also a property of behavior ... [Social forces other than
law] may give rise to patterns of behaviour that look, and in a sociological sense
are more rule-bound than behavior that is in theory rigorously structured by law"
(Lempert, 1994: 186-7). Thus, I would suggest it is, at least in principle, possible
to model legal decision-making from a perspective which accepts that legal rules
may often be peripheral to the decision environment and still be able research and
develop 'rules' (or predictable patterns) of discretionary legal practice.
If it is true, then, that legal decision-making can be both strongly characterised
by formal, legal discretion and also produce decisions that are patterned, pre-
dictable and comprehensible (Baumgartner, 1994), then any model of discretionary
decision-making on which decision-support system is based should attempt to se-
[120]
riously study informal behavioural rules. 9 I have suggested that this may uncover a
more holistic, schematic model rather than a legal-analytical one. In this sense then
sentencing is discretionary not because the judge runs out of legal rules (e.g.: Hart,
1994: 251-3), or, because s/he is unconstrained by legal precedent, but because
s/he is required to make a substantively (rather than procedurally) just decision
patterned by routine social and organisational practices.
Thus, it is not the idea itself of a computer model of decision-making based
on some notion of 'rules' that presents a difficulty, but the normally assumed
interpretation and representation of 'rules' as formal dogma implying an analyt-
ical, mathematical style of reasoning. Further research examining the character of
judicial intelligence and informal behavioural patterns is likely to help to model
the 'rules' of discretionary judicial decision-making.
I mentioned at the beginning of this subsection that the pessimism in AI & law
about the potential for modelling decision-making which is revealed as determined
by social and sociolegal rules rather than rules of 'legal' dogma is understandable. I
say this because, as Alldridge (1997) has argued, the sociolegal study 'movement'
has failed to take serious interest in work theorising the relationship between law
and computers (including applications). Such work has tended to be ignored by so-
ciolegal scholarship which has tended to regard sociolegal insights and perspectives
as unnecessary for such a 'technical' area. Given this failure to engage seriously
with thinking about computer modelling and application design, it is possible to
understand why such perspectives and insights into the nature of discretion have
barely been recognised in the work to design and theorise applications, and why
this results in (misplaced) pessimism. Yet the opportunities for enhancing the un-
derstanding of discretionary judicial decision-making, legal reasoning and how
these insights might be implemented appear to be immense. I have tried to show
in this article that research and insights from sociolegal studies and the sociology
of law can provide invaluable assistance in the task of understanding the nature
of judicial 'intelligence' and discretion. This discretion is relatively undetermined
by formal legal dogma. However, judicial discretion is revealed as patterned, pre-
dictable, and intelligible (rather than irrational, arbitrary, and chaotic) and therefore
amenable to computer modelling to produce judicial applications.
9 This may link in with the complaint that the development of legal knowledge-based applications
supporting legal practice is "rather slow". (Oskamp, Tragter, & Groendijk, 1995: 209). They suggest
that greater attention needs to be paid to studying the appropriateness of applying different strategies
to different areas of legal activities (Oskamp, Tragter, & Groendijk, 1995: 213-4). This in tum, of
course, necessitates further study and conceptualisation of the practice of legal decision processing
in its natural environment.
[ 121 ]
226 CYRUS TATA
Acknowledgements
I am grateful to the anonymous reviewers for their helpful comments. I would also
like to thank Simon Halliday, Peter Alldridge, John Wilson, and John Zeleznikow
for their valuable thoughts on an earlier draft of this paper.
References
Alschuler, A. (1991) The Failure of Sentencing Guidelines: A Plea for Less Aggregation University
of Chicago Law Review 58, 901-951.
Alldridge, P. (1997) Anoraks Among the Suits and Jeans: Computers, Law and the Le-
gal Academy. Journal of LaM\ Information and Technology 1997 (2), <http:elj. war-
wick.ac.uk/jiltlwip/97_2aldrl>.
Allen, H. (1987) Justice Unbalanced: Gender, Psychiatry, and Judicial Decisions. Open University
Press.
Armytage, L. (1995) Evaluating the Impact of Judicial Education. Journal of Judicial Administration
4(3), 143-170.
Ashworth, A. (1997) 'Sentencing by Computer: What Next?' Criminal Law Review 153-154.
Ashworth, A. (1992a) Sentencing and Criminal Justice. London: Weidenfeld and Nicholson.
Ashworth, A. (1992b) Sentencing Reform Structures. In Tonry, M. (ed.), Crime and Justice: A Review
of Research, Vol. 16. Chicago: University of Chicago Press.
Ashworth, A., Genders, E., Mansfield, G., Peay, J., and Player, E., (1984) Sentencing in the Crown
Court: Report of an Exploratory Study. University of Oxford Centre for Criminological Research,
Occasional Paper No. 10.
Bain, W. (1986) Case Based Reasoning: A Computer Model of Subjective Assessment. PhD Thesis.
Department of Computer Science, Yale University.
Bain, W. (1989) 'Judge' in Christopher K. Reisbeck and Roger C. Schank (1989) 'Inside Case-based
Reasoning', New Jersey: Lawrence Erblaum Associates.
Bainbridge, D. (1991) CASE: Computer Assisted Sentencing in Magistrates' Courts, Paper presented
to the BILETA Conference 1991.
Barnett, A., Blumstein, A., Cohen, J., Farrington, D. (1992) Not all Criminal Careers are Equally
Valid. Criminology 30(J), 133-147.
Basten, J. (1980) Judicial Accountability: A Proposal for a Judicial Commission. The Australian
Quarterly 468-485.
Basten, J. (1995) 'Should Judges Have Performance Standards?' Paper Delivered to the 1995 New
South Wales Legal Con1'ention.
Baumgartner, M.P. (1994) The Myth of Discretion. In Hawkins, K. (ed.), The Uses of Discretion.
Oxford: Oxford Socio-Legal Studies Clarendon Press.
Bench-Capon, T.M.J. (1994) In Defence of Rule-Based Representations for Legal Knowledge-Based
Systems, Law, Computers and Artificial Intelligence 1, 15-28.
Bennett, W., Feldman, M. (1981) Reconstructing Reality in the Courtroom. Tavistock Publications.
Boyle, C.K. and Allen, M.J. (1990) Sentencing in Northern Ireland. SLS Publications (NI).
Canadian Sentencing Commission (1987) Sentencing Reform: A Canadian Approach.
Chan, J. (1991) A Computerised Sentencing System for New South Wales Courts. Computer Law
and Practice 137
Chi, M. (1988) The Nature of Expertise. Lawrence Erblaum Associates.
Corkery, I.M. (1992): The Use of Vignettes in Sentencing Studies of English Magistrates. 1nteran-
tional Journal of the Sociology of Law 20, 253-270.
Council of Europe (CDPC) (1993) Consistency in Sentencing R (02) 17, October 1993, Strasbourg.
[122]
Crombag, H.P., De Wijkerslooth, J.L., and Van Tuyl Van Serooskerken, E.H. (1975) On Solving
Legal Problems. Journal of Legal Education 27, 168-201.
Curran, J. and Chambers, G. (1982): Social Enquiry Reports in Scotland. HMSO, Edinburgh.
Doob, A. (1990) Sentencing Aids: Final Report to the Donner Canadian Foundation. Centre of
Criminology University of Toronto: Toronto.
Doob, A. and Park, N. (1987) Computerised Sentencing Information for Judges: An Aid to the
Sentencing Process Criminal Law Quarterly 30, 54.
Dworkin, R (1977) Taking Rights Seriously. Harvard University Press.
Edwards, L. and Huntley, J. (1992) Creating a Civil Jurisdiction Adviser. Law, Computers and
Artificial Intelligence 1(1),5-40.
Emerson, RM. (1995) Holistic Effects in Social Control Decision-Making. In Abel, R (ed.), The
Law and Society Reader. New York: New York University Press.
Ericson, R (1981) Making Crime: A Study of Detective Work. Toronto, Butterworth.
Ericson, R. (1982) Reproducing Order: A Study of Police Patrol Work. Toronto: University of Toronto
Press.
Fitzmaurice, C. and Pease, K. (1986) The Psychology of Judicial Sentencing. Manchester:
Manchester University Press.
Fox, R. (1994) The Meaning of Proportionality in Sentencing. Melbourne University Law Reviews
19(3),489-511, at 510.
Fox, R and Freiberg, A. (1985) Sentencing: State and Federal Law in Victoria. Oxford University
Press.
Freed, ? (1992) Federal Sentencing in the Wake of Guidelines: Unacceptable Limits on the
Discretion of Sentencers. Yale Law Journal 101(8), 1681.
Giller, H. and Morris, A. (1981) "What Type of Case if This?" Social Workers' Decisions about
Children Who Offend. In Adler, M. and Asquith, S. (eds.), Discretion and Welfare. London:
Heinemenan Educational Books.
Greenberg, D. (1991) Modelling Criminal Careers. Criminology 25(1), 17-46.
Gruner, R (1991) Sentencing Advisor: An Expert Computer System for Federal Sentencing
Analyses. Paper presented at BILETA Conference.
Hart, H. (1994) The Concept of Law. Second Edition, Clarendon Press.
Hassett, P. (1993) Can Expert System Technology Contribute to Improved Bail Decisions? Interna-
tional Journal of Law and Information Technology 2.
Hawkins, K. (1992) The Use of Legal Discretion: Perspectives from Law and Social Science. In
Hawkins, K. (ed.) The Uses of Discretion. Oxford: Oxford Socio-Legal Studies Clarendon Press.
Hedderman, C. and Gelsthorpe, L., eds. (1997) Understanding the Sentencing of Women: Home
Office Research Study, p. 170. Home Office.
Henham, R. (1997) Criminal Justice and Sentencing Policy. Dartmouth.
Hogarth, J. (1971) Sentencing as a Human Process. Toronto: University of Toronto Press in
Association with the Centre of Criminology University of Toronto.
Hogarth, J. (1988) Sentencing Database System: User's Guide. Vancouver: University of British
Columbia.
Hutton N., Paterson A., Tata c., and Wilson J. (1996) A Prototype Sentencing Information System
for the High Court of Justiciary: Report of the Study of Feasibility. Edinburgh: Scottish Office
Home and Health Department, Central Research Unit.
Hutton, N., Paterson, A., Tata, c., and Wilson, J. (1995) Decision Support for Sentencing in a Com-
mon Law Jurisdiction. The Fifth International Conference on Artificial Intelligence and Law:
Proceedings of the Conference. Maryland.
Hutton, N. and Tata, C. (1995) Patterns of Custodial Sentencing in the Sheriff Court. Edinburgh:
Scottish Office Home and Health Department Central Research Unit.
Hutton, N., Tata, C. and Wilson, J. (1995) Sentencing and Information Technology: Incidental
Reform? International Journal of Law and Information Technology 2(3).
[123 ]
228 CYRUS TATA
Jareborg, N. (1994) The Sweedish Sentencing Law. European Journal on Criminal Policy and
Research 2(1).
JOrg, N., Field, S., and Chrisje, B. (1995) Are 'inquisitorial' and 'adversarial' systems converg-
ing? In Fennell, P., Harding, C., Jorg, N., and Swart, B. (eds.), Criminal Justice in Europe: A
Comparative Study. Clarendon Press.
Lawrence, J. (1988) Expertise on the Bench: Modelling Magistrates' Judicial Decision-Making.
In Chi, M., Glaser, R.. and Marshall, J. (eds.), The Nature of Expertise. Lawrence Erlburum
Associates.
Lempert, R. (1994) Discretion in a Behavioural Perpsective. In Hawkins, K. (eds.) The Uses of
Discretion. Clarendon Press. First published 1992.
Lloyd-Bostock, S. (1988) Law in Practice: Applications of Psychology to Legal Decision-Making
and Legal Skills. British Psychological SocietylRoutledge.
Lovegrove, A. (1989) Judicial Decision Making, Sentencing Policy and Numerical Guidance. New
York: Springer-Verlag.
Lovegrove, A. (1997) The Framework of Judicial Sentencing. Cambridge University Press.
Moody, S.R. and Tombs, J. (1982) Prosecution in the Public Interest. Edinburgh: Scottish Academic
Press.
Moxon, D. (1988) Sentencmg Practice in the Crown Court. Home Office Research Study No. 103,
London, HMSO.
Murbach, R. and Nonn, E. (1991) Sentencing by Artificial Intelligence Tools: Some Possibilities
and Limitations Paper presented at The Joint Meeting of the Law and Society Association and
the Research Committee of the Sociology of Law of the International Sociological Association
Amsterdam 1991.
Oskamp, A., Tragter, M., and Groendijk, C. (1995) AI and Law: What about the Future? Letter to
the Editor. Artificial Intelligence and Law 3: 209-215.
Parker, H., Sumner, M., and Jarvis, ? (1989) Unmasking the Magistrates. Milton Keynes: Open
University Press.
Parton, D.A., Hansel, M., and Stratton, J.R. (1991): Measuring Crime Seriousness: Lessons from the
National Survey of Crime Severity. The British Journal of Criminology 31, 72-85.
Potas, 1., Ash, D., Sagi, M., Cumines, S. and Marsic, N. (1998) Informing the Discretion: The
Sentencing Information System of the Judicial Commission of New South Wales. International
Journal of Law and Iriformation Technology, 6(2).
Potas, 1. (1997) Consisten,:y in Approach to Sentencing: A Description of the Judicial Commis-
sion's Sentencing Information System. Proceedings of the Judicial Decision Support Workshop.
Melbourne.
Potas, 1. (1991) The Sentencing Information System of New South Wales: Promoting Consistency in
Sentencing Through Computerisation. Paper presented to The Commonwealth Magistrates and
Judges Association Sydney 1991.
Scottish Office Home Department (June 1994) Firm and Fair: Improving the Delivery of Justice in
Scotland. H.M.S.O.
Schild, U.J. (1995) Intelligent Computer Systems for Criminal Sentencing, The Fifth International
Conference on Artificial Intelligence and Law: Proceedings of the Conference, Maryland 1995,
pp. 229-238.
Shapland, J. (1981), BeMeen Conviction and Sentence: the Process of Mitigation, Routledge &
Kegan Paul.
Shapland, J. (1987) Who Controls Sentencing? Influences on the Sentencer. In Donald C. Pennington
and Sally Lloyd-Bostock (eds.), The Psychology of Sentencing: Approaches to Consistency and
Disparity. Oxford: Centre for Socio-Legal Studies.
Simon, E. and Gaes, G. (1989) ASSYST - Computer Support for Guideline Sentencing. The Second
International Conference on Artificial Intelligence and Law: Proceedings of the Conference,
Vancouver 1989, pp. 195-200.
[ 124]
Simon, E., Gaes, G., and Rhodes, W. (1991) ASSYST - The Design and Implementation of Computer
Assisted Sentencing. Federal Probation 55, 46.
Spears, D. (1993) Providing Computerised Sentencing Information to Judicial Officers: The New
South Wales Experience. Sydney: NSW Judicial Commission.
Sudnow, D. (1964) Normal Crimes: Sociological Features of the Penal Code in a Public Defender
Office. Social Problems 12, 255-276.
Tata, C. (1998a) "Neutrality", "Choice", and "Ownership" in the Construction, Use, and Adaptation
of Judicial Decision Support Systems. International Journal of Law and Information Technology,
6(2).
Tata, C. (1998b) Problematising the "Structure" in Sentencing Discretion. Paper Presented to the
Annual Meeting of the Law and Society Association, 2-6 June 1998, Colorado.
Tata, C. (1997) Conceptions and Representations of the Sentencing Decision Process. Journal of Law
and Society 24(3),395-420.
Tata, C. and Hutton, N. (1998) What "Rules" in Sentencing? The International Journal of the
Sociology of Law, 26(3).
Tata, c., Hutton, N., Wilson, J., Paterson, A., and Hughson, I. (1997) A Sentencing Information
System for the High Court of Justiciary of Scotland: Report of the Study of the First Phase of
Implementation and Enhancement. Scottish Courts Administration.
Tata, c., Wilson, J., and Hutton, N. (1996) Representations of Knowledge and Discretionary
Decision-Making by Decision-Support Systems: the Case of Judicial Sentencing. The Journal
of Information, Law and Technology <http://elj.warwick.ac.uklelj/jiltlartifintl2tataJ>No. 2.
Tata, c., Hutton, N., Wilson, 1., Paterson, A., and Hughson, I. (1998) A Sentencing Information
System for the High Court of Justiciary of Scotland: Report of the Study of the First Phase of
Implementation and Enhancement. (Scottish Courts Administration).
Tito, C. (1987) Artificial Intelligence: Can Computers Understand Why Two Legal Cases are
Similar? Computer/Law Journal 7, 409-437.
Tonry, M. (1987) Sentencing Guidelines and Sentencing Commissions: The Second Generation. In
Wasik and Pease (eds.) Sentencing Reform: Guidance or Guidelines? Manchester: Manchester
University Press.
Tonry, M. (1992) Judges and Sentencing Policy - The American Experience. In Munro and Wasik
(eds.), Sentencing, Judicial Discretion and Training. London: Sweet and Maxwell.
Tonry, M. (1993) The Failure of the U.S. Sentencing Commission's Guidelines. Crime and
Delinquency 131.
Tonry, M. (1996) Sentencing Matters. Oxford: Oxford University Press.
Van Duyne, P. (1987): Simple Decision Making. In Pennington, D.C. and Lloyd-Bostock, S. (eds.),
The Psychology of Sentencing: Approaches to Consistency and Disparity. Oxford: Centre for
Socio-Legal Studies.
Van Noortwijk, K. and De Mulder, R. (1997) The Similarities of Text Documents. Journal of
Information, Law, and Technology http://jilt.law.strath.ac.ukljiltlartifintl97_2noor/noor.htm.
Von Hirsch, A. (1976) Doing Justice: The Choice of Punishments. New York: Hill and Wang.
Von Hirsch, A. (1993) Censure and Sanctions. Oxford: Clarendon Press.
Von Hirsch, A. and Jareborg, N. (1989) Sweden's Sentencing Statute Enacted. Criminal Law Review
275.
Warner, D. (1990) The Role of Neural Networks in Law Machine Development. Rutgers Computers
and Technology Law Journal 16: 129-144.
Wilkins, L., Kress, 1., Gottfredson, D., Calpin, J., and Gelman, A. (February 1978) Structur-
ing Guidelines: Structuring Judicial Discretion. Report on the Feasibility Study. U.S. Justice
Department.
Zdenkowski, G. (1986) Sentencing: Problems and Responsibility. In Chappell, D. and Wilson, P.
(eds.) Australian Criminal Justice System. Butterworths.
[ 125]
230 CYRUSTATA
Zeleznikow, J. and Hunter, D. (1994) Building Intelligent Legal Iriformation Systems. Computer Law
Series 13, Kluwer Law and Taxation Publishers.
Zeleznikow, J. Stranieri, A, ,md Gawler, M. (1996) Project Report: Split-Up - A Legal Expert System
which Determines Property Division upon Divorce. Artificial Intelligence and Law 3, 267-275
Zeleznikow, J. and Stranieri, A. (1997) Modelling Discretion in the Split-Up System. Proceedings
of the Workshop on Judicial Decision Support Systems, The Sixth International Conference on
Artificial Intelligence and Law. Melbourne.
[ 126]
231
Modelling Reasoning with Precedents in a Formal

Dialogue Game
HENRY PRAKKEN*
Department of Mathematics and Computer Science, Free University Amsterdam, The Netherlands.
E-mail: [email protected], http://www.cs.vu.nOhenry
GIOVANNI SARTOR
Faculty of Law, The Queen's University of Belfast, Belfast BT7 1NN, Northern Ireland.
E-mail: [email protected], http://www.qub.ac.uklgslsartorl.htm
Abstract. This paper analyses legal reasoning with precedents in the setting of a formally defined
dialogue game. After giving a legal-theoretical account of judicial reasoning with precedents, a for-
mal method is proposed for representing precedents and it is discussed how such representations can
be used in a formally defined dialectical protocol for dispute. The basic ideas are to represent cases
as argument structures (including pro and con arguments, and the arguments for adjudicating their
conflicts) and to define certain case-based reasoning moves as strategies for introducing informa-
tion into a dispute. In particular, analogizing and distinguishing are conceived as elementary theory
construction moves, which produce new information on the basis of an existing stock of cases. The
approach also offers the possibility of using portions of precedents and of expressing criteria for
determining the outcome of precedent-based disputes.
The analysis, which is partly based on argument-based semantics of defeasible reasoning, has two
aims. The first is to provide a formalization of certain aspects of legal theories on judicial reasoning
and judge-made law, and the second is to provide formal foundations for certain aspects of computer
programs for case-based reasoning in the legal domain.
Key words: Defeasible argumentation, case-based reasoning, precedents, normative dialectics
1. Introduction
General Setting
Research on case-based reasoning (CBR) is one on the main streams of AI &
law (e.g. McCarty and Sridharan, 1981; Rissland and Ashley, 1987; Ashley, 1990;
Berman and Hafner, 1991; Skalak and Rissland, 1992; Branting, 1994 and Aleven
* Much of the research reported in this article was done while the first author was employed
at the ComputerlLaw Institute of the Faculty of Law, Free University Amsterdam, supported by a
research fellowship of the Royal Netherlands Academy of Arts and Sciences, and partly by Esprit
WG 8319 'ModeIAge'. The final version was written while the first author was employed at the
Institute of Applied Information Technology of the GMD Bonn, Germany, supported by VIM (A
VIrtual Multicomputer), a project funded by the EC's Human Capital and Mobility programme.
[ 127]
232 HENRY PRAKKEN AND GIOVANNI SATOR
and Ashley, 1996). This research has provided not only computer applications, but
also models and insights relevant for the theoretical understanding of judge-made
law, which parallel the investigations of legal theory (e.g. MacCormick, 1978;
Goldstein, 1987; Raz, 1989 and Cross and Harris, 1991). In particular, it has fo-
cused on the dialectical process of citing and comparing cases, and on the various
heuristics of case-based reasoning.
Another development in AI & Law is logical research on nonmonotonic, or de-
feasible legal reasoning (e.g. Sartor, 1992; Prakken, 1993; Gordon, 1995; Prakken
and Sartor, 1996b; Verheij, 1996 and Hage, 1996, 1997). Here the main concern is
to give logical accounts of legal reasoning with incomplete, uncertain or inconsis-
tent knowledge. This development draws on and adds to the tools of nonmonotonic
logic.
A particularly useful tool has been found in logical systems for defeasible argu-
mentation, which model nonmonotonic reasoning as the construction and compari-
son of (logical) arguments for and against a certain proposition (e.g. Pollock, 1987;
Loui, 1987; Dung, 1995; Vreeswijk, 1997). In our opinion, these systems provide a
tool for connecting and integrating the two research developments just mentioned
since, unlike other nonmonotonic logics, they do justice to the dialectical structure
of case-based knowledge and to the adversarial procedure of case-based reasoning.
In particular, they make it possible to model case-based reasoning as a special type
of defeasible argumentation, which combines a logical system (in the tradition of
nonmonotonic logics) and a specific set of argument moves and heuristic strategies
(in the tradition of case-based research). The first contributions adopting this ap-
proach were of Ron Loui and his colleagues (Loui et aI., 1993; Loui and Norman,
1995), and the present paper (which is a revised and extended version of Prakken
and Sartor, 1997b) further develops the attempt of modelling case-based reasoning
with the help of logical argumentation systems.
Focus of Research
In pursuing our aim, we focus on reasoning with precedents in an adversarial set-

ting. We first give a legal-theoretical account of judicial precedent-based reasoning,
resulting in a set of requirements for formal and computational models. Then we
present a formal model of dialectical reasoning with precedents that aims to satisfy
these requirements, and that is defined on top of a logical system for defeasible
argumentation. This system is the one we previously developed in Prakken and
Sartor (l996b; 1997a), and which is based on the abstract logical approach to
defeasible argumentation of Dung (1995) and Bondarenko et aI. (1997). We then
use our formal model in an analysis of aspects of computer programs for case-based
legal reasoning. It is this analysis which explores the connection between the two
research developments. At the same time, our formal model can be regarded as a
formalization of the discussed aspects of legal theories on judicial reasoning, and
therefore as a contribution to legal theory.
[ 128]
MODELLING REASONING WITH PRECEDENTS IN A FORMAL DIALOGUE GAME 233
We shall in particular focus on HYPO style case-based reasoning, i.e., on the

dialectical interchange of arguments which support or oppose a claim by citing,
analogizing or distinguishing legal precedents (Rissland and Ashley, 1987; Ash-
ley, 1990). Our aim here is twofold. Firstly, we want to model some of HYPO's
argument moves within our formal dialogue game for reasoning with conflicting
arguments of Prakken and Sartor (l996b; 1997a). And secondly, we want to pro-
pose a richer method for representing cases than is allowed by HYPO. We shall
also briefly compare our proposal to some other extensions of and alternatives to
the HYPO approach, the CABARET system of Skalak and Rissland (1991), the
work of Branting (1991; 1994), and the CATO system of Aleven and Ashley (1996;
1997).
Our proposal has the following main ingredients. Both cases and case-based
reasoning are seen as pieces of argumentation. Firstly, each past case is represented
as a completed or frozen piece of argumentation, i.e., as a dialectical argument
structure. Such a piece of argumentation consists of a set of premises ground-
ing (possibly) conflicting arguments, where the winning argument supports the
decision of the case. Secondly, case-based reasoning concerning a new situation
will in our approach be modelled as a process of argumentation, where each argu-
ment move may make direct or analogical use of precedents, or may distinguish
precedents used by the opponent. What is also important is that the dialectical
interchange of arguments will conform to the rules of the dialogue game defined
in (Prakken and Sartor, 1996b) but that, unlike there, the parties will be free to
introduce new premises during the dialogue. In fact, one of our main aims is to
reconstruct case-based reasoning moves as rational heuristics for introducing new
premises into a dispute. In this way we hope to reproduce the basic forms of
precedent-based reasoning (following a precedent, analogizing it, distinguishing
it from the present case) while embedding such reasoning forms in a more general
formal model of defeasible argumentation. It should be stressed that our model,
although choosing for a particular type of analogical reasoning, does not crucially
depend on this choice; it is compatible with diverse approaches to analogy and
theory construction in case-based reasoning.
Nature of our Research

There are two important differences between our model and the other systems that
we shall discuss. The first is that while HYPO, CATO, CABARET and Branti-
ng's GREBE system are implemented systems, we present a more abstract, logical
model. We think that such abstract models are a useful complement to the develop-
ment of actual systems and computational models. A logical formalization makes it
possible not only to disambiguate and make things precise, but also to prove formal
properties, to view clearly similarities and differences between various systems and
approaches and to assess the possibility of their integration. Furthermore, it may
show how results obtained in other areas (for instance, proof theory), can be made
available.
[129]
A second difference is that while HYPO, CABARET and CATO have a more
'cognitive' approach, aimed at generating realistic disputes, our approach is more
'normative', defining a rational procedure for testing the tenability of a claim. This
requires some explanation. Our aim is to apply the traditional dialectical method
as studied by philosophers, which is aimed at testing the tenability of a claim in
a dialectical inquiry (see e.g. Rescher (1977) for an application of this method
in epistemology, and Loui (1998) for a defence of this method in nonmonotonic
reasoning). The adoption of a normative approach has two important consequences.
The first is what we call 'dialectical asymmetry'. The proponent and the op-
ponent of a claim have different tasks: the proponent must prove that the claim is
tenable, while the opponent just has to prevent the proponent from doing so; it is
not the opponent's task to prove that the opposite claim is tenable. The second con-
sequence of normative dialectics is that it must be possible to determine the relative
strength of each move, to see whether it adequately responds to the other party's
previous move. In particular, while the proponent's arguments must be stronger
than the opponent's previous move, the opponent's arguments only have to be not
weaker than the proponent's previous move.
In CBR systems these normative aspects are largely absent: for instance, the
intended output of HYPO and CABARET is not an answer to the question whether
a claim is tenable; instead, the intended output is a dispute as it could take place
between 'good' lawyers. It should be noted, however, that the difference is not
clear-cut. We are also interested in defining realistic disputes; our model can be
said to define which of those disputes conform to the ideal of normative dialectics.
And HYPO and CABARET also have criteria for the strength of arguments: for
instance, each cited case must be as similar as possible to the current fact situation.
This rule (and others) can be said to prune the space of possible disputes. But
HYPO and CABARET have no dialectical asymmetry: the rules are the same for
the plaintiff and the defendant. And these systems do not implement a notion of
'winning' a dispute (except in a few cases). Instead, they assume that the final
choice is made outside the system.
Background: Four Layers in Legal Argument
We shall carry out our investigation against the background of a four-layered pic-
ture oflegal argumentation, discussed in (Prakken, 1997) and (Sartor, 1997).1 The
first layer (the logical one) provides the logical structure of single arguments, i.e.,
it defines how pieces of information can be combined in order to provide basic
support for a claim. The second layer (the dialectical one) focuses on conflicting
arguments: it introduces such notions as 'counterargument', 'attack', 'rebuttal' and
'defeat', and it defines how, given a set of premises and evaluation criteria, it can
1 The first three layers were earlier distinguished by Prakken (1995) and Gordon and Kara-
capilidis (1997), while a somewhat different three-layered model was proposed by Brewka and
Gordon (1994).
[130 ]
be detennined which of the possible arguments prevail. These are the notions de-
fined by the above-mentioned logical systems for defeasible argumentation. And
since Dung (1995) has shown that (more or less) any nonmonotonic logic can be
reformulated as such a system, one can say that the dialectical layer is the layer
that is addressed by nonmonotonic logic.
The third layer (the procedural one) regulates how an actual dispute can be
conducted, i.e., how parties can introduce or challenge new information and state
new arguments. In other words, this level defines the possible speech acts, and
the discourse rules for when and how these speech acts can be performed. Thus
the procedural layer differs from the first two in one crucial respect. While the
logical and dialectical layer assume a fixed set of premises, at the procedural layer
the set of premises is constructed dynamically, during a debate. This also holds
for the final layer, the strategic or heuristic one, which provides rational ways of
conducting a dispute within the procedural bounds stated at the third level; i.e.,
it concerns heuristics and strategies for expanding the available knowledge and
constructing new theories.
All four layers are to be integrated into a comprehensive view of argumentation:
the logical layer defines, by providing a notion of arguments, the objects to be
evaluated at the dialectical layer; the dialectical layer offers to the procedural and
heuristic layers a judgement of whether a new argument might be relevant in the
dispute; the procedural layer constrains the ways in which new inputs, supplied by
the heuristic layer can be submitted to the dialectical one; the heuristic layer pro-
vides the matter which is to be processed in the system. Each layer can obviously
be studied (and implemented) in abstraction from the other ones. For example, the
study of the dialectical layer can abstract from the procedural and the heuristic
layers, when the pool of given information is fixed; the study of the procedural
layer can abstract from the heuristic level when ways for regulating the interaction
of human beings are considered; the study of the heuristic layer can just focus on
ampliative strategies, regardless of their procedural admissibility and of the dialec-
tical evaluation of their products. However, it would be a grave misconception,
and a serious hindrance to the development of a large-scope formal theory of legal
reasoning, to regard those layers as alternative rather than as complementary.
In our opinion, much AI & law research can be classified into this four-layered
model of argumentation. The logical layer includes logical deduction, and the ba-
sic reasoning forms of rule-based expert systems (forward chaining and backward
chaining). The dialectical layer was addressed by our previous work, which defines
the status of arguments on the basis of a given pool of conflicting premises. The
procedural level is the central focus of Gordon's (1995) Pleading Game, which
studies argumentation protocols for certain types of legal disputes, and of Hage et
al. (1994), who give a procedural account of hard cases. The fourth level, finally,
has been especially studied in relation to case-based reasoning, as modelled in e.g.
[ 131 ]
HYPO, CATO, CABARET, Branting's work, Loui and Norman (1995), and by
McCarty and Sridharan (1981) and McCarty (1995).1
The present paper also addresses the fourth layer of argumentation, in the con-
text of case based reasoning. In particular, we regard HYPO-style analogizing and
distinguishing as heuristics for adding new information into a dispute, formalizing
the view which was earlier defended in (Prakken, 1995). As far as the first and
second layer are concerned, we shall build upon our formalization in (Prakken and
Sartor, 1996b, 1997a). In the present paper we abstract from the third level, apart
from one comment in Section 5.4. on the issue of how a precedent can be said to
control a judicial decision.
Outline afthe Paper

We start our investigations in Section 2 with a legal-theoretical account of judicial
precedent-based reasoning, resulting in a set of requirements for formal and com-
putational models. Then, in Section 3, we briefly discuss some (computational)
models of case-based reasoning, especially in regard to those requirements, and
we discuss the need for extending these models.
In Section 4 we present the building blocks for our proposal: our previously
developed logic for defeasible argumentation, and a new method for representing
precedents. Our formal model of precedent-based legal reasoning is then presented
in Section 5, after which it is applied to an extended example in Section 6 and
compared to the earlier-discussed systems in Section 7.
2. Legal-theoretical Considerations on Precedent-based Judicial Reasoning

In this section we give a legal-theoretical account of judicial precedent-based rea-
soning. In particular, we analyse the internal structure of a precedent (2.1) and the
role of a precedent in further legal decision making (2.2). In this section it is not
our aim to be original: instead we want to provide an analysis of some aspects of
the doctrine of precedent which can serve as the basis for the formal part of this
article and for comparison with other relevant work. Another aim of this overview
is to show that many observations in the AI & law literature have their counterpart
in legal theory and jurisprudence.
2.1. A DIALECTICAL ACCOUNT OF JUDICIAL OPINIONS

The expression 'case' is, as it is well known, an ambiguous one, even in legal
contexts. It may refer to the whole proceedings of a lawsuit, but more frequently
it just refers to the final act of those proceedings, i.e., the decision of the judge
supported by his/her opinion. Only in this second meaning can cases usually be
1 In our opinion, the third and fourth layer address what McCarty (1997) calls the theory
fonnation aspect of legal argument.
[ 132]
found in law reports and can they be said to constitute an authority for future
decisions and to be relevant for case-based reasoning.
We claim that in general cases have a dialectical structure, i.e., they contain
not only arguments supporting the decision but also arguments attacking it, and
arguments why these attacks do not succeed. However, in considering whether
cases have a dialectical structure, the two notions of 'case' just sketched must
be distinguished. The assertion that legal proceedings normally do (and should)
exhibit a clear dialectical structure is just a truism: they exemplify in the most clear
and paradigmatic way the dialectical process of disputation. This process starts
when the plaintiff and the defendant identify the problem and produce conflicting
arguments for its alternative solutions and it terminates when the judge determines
the output of the proceedings. The assertion, instead, that judges' opinions are
'dialectical' requires further considerations and qualifications.
We argue that, although being a monological discourse, often a judge's opinion
reproduces (and possibly extends) the dialectical context ofthe disputation between
the parties. This is the case when the judge explicitly answers the issues raised
in the parties' debate, on the basis of an evaluation of their arguments (and of
the further arguments provided by the judge himlherself). Legal justification, at
least in hard cases, requires that "so important an aid to the intelligent and living
apprehension of a truth, as is afforded by the necessity of explaining it to, and
defending it against, opponents" (Mill, 1974). Therefore, at least when significant
legal problems have to be solved, a judge's opinion may present three features,
which should be preserved in the formal and computable representation of the
precedent (when practically feasible).
Firstly, the opinion may contain not only the argument supporting the adopted
decision, but sometimes also the defeated arguments to the contrary. This is since
in such cases a justification limited to the winning argument would not express an
adequate rationale and would fail to clarify the context and the limits of the winning
argument. Secondly, the opinion may contain complex arguments, which require a
sequence of steps before reaching the desired conclusion; and each of those steps
may involve the necessity of adjudicating the conflicts with arguments to the con-
trary. Preliminary or prejudicial questions need to be solved in order to tackle the
final substantial point (or an argument pleading for the substantial point may be
challenged by raising issues concerning preliminary points). A representation that
compresses judicial reasoning in the immediate connection between facts of the
case and final decision misses a fundamental feature of judges' decision making.
And finally, the opinion may include more argument-layers: the conflict between
basic arguments needs to be adjudicated by higher level arguments, which may
again be in conflict, etc. Without those higher level arguments, which substantiate
a rationale for decisions on controversial points, judicial reasoning would in some
cases appear impoverished and arbitrary.
In conclusion, we argue that a satisfactory model of precedents should account
for the possibility of representing cases as sets of (possibly) conflicting, multi-
[ 133 ]
step and multi-level arguments. Precedents just including consistent, one-step and
one-level arguments should be considered as limiting cases of a richer framework.
Obviously, this possibility is not intended as a facility to be offered by each com-
putable formalization of case-based reasoning (and each computer application).
It is only a requirement for a dialectical theory of precedents, which can be ade-
quately constrained in specific application domains. Let us now examine the three
'internal' features of judicial precedents in more detail.
2.1.1. Judicial Rationales as Dialectical Multi-argument Structures

The traditional view of the justification of legal decisions is the deductive one:
to justify a decision means to produce a consistent set of legally valid and/or
factually true premises which logically imply the decision of the case. Differ-
ent variants of the deductive model can be found, according to the nature of the
premises and legal sources from which those premises have to be obtained. The
main alternative lies in the distinction between legalistic approaches (according to
which those premises are to be found in law texts) and conceptualist approaches
(according to which those premises are to be found in jurists' definitions). Nev-
ertheless, the deductive approach is also compatible with other ways of obtaining
the premises (from precedents, from a natural law code, from social customs, etc.).
Some authors have therefore affirmed that the need of a deductive justification for
judicial decision making is independent from the nature of the premises of such a
deduction. In their opinion, the advantage of a deductive justification consists in
making those premises explicit, and so in facilitating control and critique (Klug,
1966). Nowadays this neutral version of deductivism seems to be accepted by
most legal theorists, who frequently stress the necessity that every legal decision is
given a deductive justification, frequently also called the internal justification (cf.
MacCormick, 1978; Wroblewski, 1983 and Alexy, 1989, pp. 221-230).
However, we argue that this 'neutral' version of deductivism is also inadequate,
because of the disputational nature of legal reasoning and particularly of judicial
reasoning. To justify a judicial decision it is not always sufficient to produce a
single argument; sometimes it is necessary to establish that the winning argument
prevails over all arguments to the contrary, at least when those arguments were pre-
sented by the losing party. Defeated arguments are also fundamental for qualifying
the strength of the victorious thesis in future cases.
Let us consider, for example, the Donoghue v. Stevenson case ([1932] AC 562),
a bench-mark case of English tort law, in which a manufacturer was held respon-
sible for marketing a bottled ginger ale containing a snail, on the basis of the rule
that marketing a defective product determines the liability of the manufacturer (we
simplify the original rule for clarifying the example). In their opinion, the judges
disposed of certain counterarguments, such as the counterargument that no remedy
should be available if no contractual relation exists between the manufacturer and
the consumer. Such counterarguments cannot therefore be successfully produced
in future cases (unless the authority of the precedent is questioned). However, this
[134 ]
does not hold for those counterarguments which are not adduced in the precedent,
and especially for those that were not grounded in the facts of the precedent (for
instance, the arguments that no remedy should be available because of a disclaimer
by the manufacturer, or because of the knowledge of the defect by the consumer).
Those counterarguments can be accepted in new cases as distinctions that restrict
the ratio of the precedent, without questioning its authority, as we shall see in the
following.
In this perspective, judicial reasoning seems to consist of an exercise in "unilat-
eral dialectic", intended as a disputational model of inquiry in which "one develops
a thesis against its rivals, with the aim of refining its formulation, uncovering its
basis ofrational support, and assessing its relative weight" (Rescher, 1977, p. 47).
Such a dialectical exercise is sometimes required not only for solving the case,
but also for building a rationale for its decision, intended as "an architectonic ally
organised structure of contentions and grounds" (Rescher, 1977, p. 53). Such a
rationale should contain plausible pro and con arguments on relevant issues, and
the evaluation of their respective strength.
A dialectical style in judges' opinions is specifically promoted by the dialog-
ical function of a judge's rationale: the judge must reply to the arguments of the
parties, and particularly to those coming from the losing one. The degree in which
opinions are in fact dialectical varies in different legal cultures. In particular, ap-
pellate decisions in common law usually exhibit a more elaborate argumentative
mode. Nevertheless, civil law judges are sometimes also explicitly urged to argu-
mentatively justify their choices. For example, Taruffo (1975, p. 266) qualifies as
fictitious those justifications that "whenever a statement represents an hypothesis
chosen by the judge within a range of alternatives" present this statement as the
only possible solution to an issue, without justifying its choice. In this perspec-
tive, those opinions making no mention of plausible contrary arguments are to be
qualified as legally defective, lacking a sufficient justification.
2.1.2. Judicial Arguments as Multi-step Structures

Judicial reasoning frequently proceeds in a stepwise manner. Generally, judicial
arguments consist of a sequence of linked inference steps. In each step certain
conditions support a certain conclusion, according to a certain general inference
rule. In a stepwise argument, intermediate conclusions become the preconditions of
further inferences until the final conclusion is reached. For example, in cases con-
cerning tort responsibility, preliminary rulings may be stated on questions such as
the negligence of the defendant or the existence of a causality relationship between
herlhis behaviour and the damage. Moreover, in cases concerning the vicarious
liability of the employer for the harm caused by the tort of the employee, a ruling
may be necessary as to the existence of an employment relation.
We argue that such a stepwise structure should be reflected in the representation
of the precedent, which should reproduce the dialectics of preliminary decisions,
without compressing all reasoning into just one step. Such a compression in fact
[ 135 ]
means a loss of information and limits the possibility of extracting rules relevant
for other cases (likewise Branting, 1994). So, for example, the ruling on the notion
of employment established in cases concerning employer's liability is prima facie
relevant for subsequent cases concerning social security duties of the employer.
Obviously, such an application of a rule outside its original argumentative context
is highly defeasible: in some cases the functional relation of a preliminary ruling
to the final decision may exclude its application to a different type of case (for
example, different areas of law or different factual contexts may require different
notions of negligence). Nevertheless, in the absence of information to the contrary,
all rules in the stepwise argument leading to the solution of the case are to be
considered autonomous case-law rules, transferable to new cases.
This view subsumes those legal theories which stress the possibility of hav-
ing multiple rationes in a case, when the judge, in order to justify its decision
has to settle multiple different points of law, or when he/she provides alternative
justifications (MacCormick, 1987).
2.1.3. Judicial Rationales as Multi-level Structures

Finally, we consider whether the solution of a conflict between two arguments
is always an unreasoned decision pertaining to the free evaluation (the sovereign
prerogative of choice) of the judge, or whether it may need a justification by means
of higher-level arguments. In this respect we agree that, sooner or later, reasoning
must stop, having reached a point such that (according to the judge's view) no
further plausible reasons can be found to question a plausible argument. This stop-
ping, however, does not need to happen immediately after noticing that conflicting
rulings apply to the case. In many cases the judge may consider (and one party may
present) a rationale for hislher choice for one of the conflicting arguments.
This aspect is usually tackled by legal theorists under the label of second-order
or external justification (cf. MacCormick, 1978, pp. 101 ff. and Alexy, 1989, pp.
230 ff.). In our perspective, such second-order justification involves two different
aspects: producing arguments which substantiate or back a questioned premise
(i.e., which conclude that the premise is applicable); and producing arguments
which adjudicate conflicts between other arguments.
Here we focus on the second aspect (for analyses of arguments of the first type
see e.g. Hage, 1996, 1997). If the judge has the duty of producing a convincing
rationale, then whenever presented with a really controvertible conflict of argu-
ments he/she should provide not only a preference, but also convincing reasons for
this preference. Those preference reasons are also important for determining the
strength of the winning argument and the possibility of successfully extending it
to new cases: if the conditions which allow to adjudicate the conflict in favour of a
certain argument do not hold in the new case, then the adjudication of that conflict
can rightly be questioned. On the other hand, preference reasons having a general
character should also be extensible to new cases, in order to adjudicate those ar-
gument conflicts which are subsumable under them. This seems to correspond to a
[136]
widespread judicial practice, which uses preference reasons in new cases, although
those cases do not directly concern the solution of the same substantial issue. Con-
sider for example the Simmenthal case, a famous decision of the European court
of Justice (1978 ECR 777) where EC law was said to prevail over national laws of
the member states, in order to solve a specific conflict between a national law rule
and a European one. The Simmenthal ruling was later used in a number of cases in
order to support the preference of other European rules against other national ones.
2.2. THE BINDING CONTENT AND THE DYNAMICS OF CASE-BASED LAW

So far we have focused on the content of an individual decision. Let us now look
at the dynamic features of judicial precedent-based reasoning, i.e., how does a
precedent affect decisions in new cases? Different views exist on the theoretical
definition of the ways in which precedents affect future decision-making. In this
debate at least three orders of problems are addressed:
1. Providing the structure of the basic dialectical argument moves (when can a
precedent be directly followed, how can it be analogized to a new situation,
how can analogy be countered by distinguishing?).
2. Identifying and deciding the conflicts of precedents (when does a precedent
govern the case, and what should be done with conflicting precedents?);
3. Understanding the dynamics of case law (how can a precedent have a fixed
content, if its relevance is to be continuously reassessed in the framework of
the subsequent decisions?).
2.2.1. Argument Moves in Reasoning with Precedents

Let us first consider the basic ways of using precedents in solving new cases.
Clearly, if precedents have a dialectical structure, then the solution of new cases
also proceeds dialectically. Therefore the use of precedents for solving a new case
should be modelled as a dialectical process. In such a process at least the following
three types of moves should be possible, in regard to a precedent: following a
precedent, analogizing it, and distinguishing it.
Before considering these moves, let us remark that according to the features de-
scribed above, precedents may have an elaborate argumentative architecture: they
include multi-step arguments where certain facts support, according to a certain
rule,3 a certain legal qualification, which in turn, according to another rule, may
support a further legal qualification, and those arguments may be opposed and
compared. It is not the whole precedent that is mentioned in those moves, but
just the relevant portions of it, i.e., the rules which may be useful for supporting
the party's contention. Let us also remark that a precedent can be cited not only
3 By a rule we mean any general warrant (or inference policy) according to which a certain factor
defeasibly supports a certain conclusion. More specific notions of rules, such as those assumed when
(rigid) rules are opposed to principles (Dworkin, 1977), or when (exclusionary) rules are opposed to
first order reasons (Raz, 1975) are not considered in our framework.
[137 ]
in order to obtain the same decision in the current case, but also to support an
outcome opposite to the one that the precedent actually had (a possibility that
was suggested by Berman and Hafner, 1991, p. l7). This happens when in the
precedent the judges stated a rule although this was successfully countered by an
exception: such a rule can take the lead in subsequent cases where the exception
does not hold. In common law systems this is the standard way of citing the famous
case of Hedley Byrne & Co Ltd v Heller and Partners (1963) in which the British
House of Lords stated that as a general rule all professionals are responsible for
their negligent statements even if the damage does not concern the client, but then
acquitted a professional who expressed a negligent statement since he had made a
disclaimer. This precedent is usually cited in order to support the responsibility of
a professional when no such disclaimer is made, that is, for the opposite outcome.
For these reasons our notion of case rules stretches beyond the usual definition
of a ratio decidendi in that it also includes those rulings contained in losing ar-
guments. In fact, if the judge felt the need to consider those rulings, this means
that they have a certain strength, so that they might take the lead in other cases,
in which prevailing reasons to the contrary do not apply. An important point is
that the possibility of citing each rule in a precedent requires ways of establishing
the rule's strength which go beyond the dichotomy of ratio decidendi and obiter
dicta. In this regard, we agree with MacCormick (1987) that each precedent ruling
is only binding "relatively to the cases and the arguments put by given parties".
Accordingly, we shall provide a general flexible mechanism for establishing the
strength of any ruling contained in a precedent, according to its argumentative
function.
Following a Precedent
The first argument move, following a precedent, applies when the mentioned prece-
dents directly governs the new case. From our perspective, following a precedent
just consists in selecting a rule from the available precedents and directly using
it in an argument concerning the new case. For instance, recall the Donoghue v.
Stevenson case, where a manufacturer was held responsible for marketing a bottled
ginger ale containing a snail, and assume that a new case comes up, where an
adulterated bottle of wine is marketed: then the Donoghue rule that marketing a
defective product determines the liability of the manufacturer can directly support
the liability of the wine manufacturer in the new case.
Analogizing a Precedent
The second argument move, analogizing a precedent, consists in using the prece-
dent in order to support the same decision a new case, although the new case is not
directly governed by the precedent. From our perspective, this basically consists
in producing a new rule (the analogy) which covers the new case, by modifying
a precedent rule which cannot be directly applied to the new case. In the present
[ 138]
paper we consider the simplest of such modifications, which consists in broadening

a rule from the precedent, i.e., in cancelling one or more of the conjunctive condi-
tions (factors) required by that rule (since those conditions are not satisfied in the
current case).
However, many other ways of producing analogies exist, for instance, those
based on abstraction, where a factor, instead of being cancelled, is replaced by a
more abstract concept, in such a way that both the replaced factor from the prece-
dent and a new factor in the current case are an instance of this abstract concept. In
Dutch civil law, a classic example of this type of analogl concerns the analogical
application of a statutory rule, viz. Section 1612 of the Dutch Civil Code. This
rule says that selling living accommodation does not affect an existing lease. This
rule was analogically applied to a case where a house was not sold but donated,
by arguing that Section 1612 is based on the principle that no transfer of property
affects an existing lease, and by then observing that both selling and donating are
instances of transfer of property. Thus the rule was first broadened by replacing
'selling' with the more abstract concept 'transfer of property', after which the rule
resulting from the broadening (the 'broadening rule') was directly applied to the
new case.
Distinguishing a Precedent
The third argument move which we consider is distinguishing, which consists in

arguing that the new case is different from the precedent, and must therefore be
differently decided, without implying that the precedent was wrong. This may
happen in quite different ways. Legal doctrine (e.g. Williams, 1982) differentiates
non-restrictive and restrictive distinguishing: the first takes place when the rejected
precedent's rule does not directly cover the new case, the second when it does but
the new case has a new element, not present in the precedent.
Non-restrictive distinguishing takes place when an analogy is contested. In our
reductive view on analogies as broadenings, this form of distinguishing consists
in arguing that some of the conditions which were cancelled (in producing the
broadening) from the original rule are essential for supporting the conclusion of
that rule, and that therefore, unless that condition holds, that conclusion cannot
be derived. Such an attack may fail if further reasons are proposed showing that
the consequent of the precedent should hold even in the absence of the missing
element. For example, the analogy drawn in Haseldine vs. Daw ([1941] 2KB 343)
was distinguished by pointing to the fact that repairers are not manufacturers, as
required by the original Donoghue rule. However, the judges rejected the distinc-
tion, accepting that the basic reasons supporting the liability of the producer would
also support the responsibility of providers of services, such as repairers.
4 Earlier discussed in Prakken (1993), pp. 22-3. Another example is discussed by Branting and
Porter (1991), as an instance of what they call "case elaboration".
[139]
Let us now move to restrictive distinguishing. Here the original rule directly
covers the new case, but it is argued that the new case has an additional feature
which impedes us from drawing the conclusion established by that rule. In other
words, restrictive distinguishing consists, from our point of view, in attacking a
precedent-based argument by means of a convenient counterargument (based on
the features of the new case). As an (unsuccessful) example, consider the case
Grant v. The Australian Knitting Mill ([1936] AC 85), where a customer had con-
tracted dermatitis from wearing pants containing an excess of sulphites, and cited
Donoghue. The defendant tried to distinguish by affirming that the pants where in
a paper envelope, supposed to be open, so that there was the theoretical possibility
of discovering the defect before the sale. The judges however, refused to accept the
relevance of this aspect since in both cases the article was supposed to reach the
consumer or user subject to the same defect as it had when it left the manufacturer.
A successful counterexample was put forward in the case Fair v. Butters ([1932]
AC 562) where a workman died because of a defective crane, and compensation
was asked citing Donoghue. Here the defendant was successful in distinguish-
ing, since the fact that the workman knew the defect (since he had put the crane
together) was considered sufficient to exclude liability of the manufacturer.
Overruling a Precedent
Distinguishing in a proper sense is impossible when the factor which (arguably)
pushes the decision in the opposite direction was already included in the precedent.
In such a context, we can no more say that the precedent's decision was right,
although the new case can be distinguished from it. The precedent already com-
pared the argument leading to the precedent's decision and the contrary argument
based on the factor we are pointing to, giving precedence to the first one. If we
do not accept this evaluation we must say that the precedent was wrong, so that
an adequate solution in the new case requires overruling it. From our point of
view, overruling cons.ist in defeating precedent-based arguments on the basis of
a counterargument based on substantive considerations.
2.2.2. Conflicting Precedents

It frequently happens. that different precedents are analogous to a new problem
situation and that those precedents point to opposed outcomes, so that a choice
must be made. One criterion for making the choice is which precedent is more
similar to the new case. In AI & Law this criterion has been extensively studied
(cf. HYPO's 'more-on point' relation). However, we like to stress that besides
similarity lawyers may also use other criteria. As Summers (1997, p. 53) affirms,
when faced with a conflict of precedents, "courts have a variety of methods" by
which to solve conflicts of precedents: first "the court should determine whether
one precedents comes from a court higher in the judicial hierarchy", alternatively,
it could "take note of which of the precedents in conflict is more recent" or "choose
[ 140]
the precedent that seems best justified in substantive policy" or "that seems to
do justice in the case under consideration". Note that the hierarchical and the
recency criterion are also reflected by the well-known common law doctrine of
(implied) overruling, when the subsequent Court has such a power; cf. for English
law, R. v. Porter [1949] 2KB 128 at 132, cited by Cross and Harris (1991, pp.
132 ff.). See for temporal considerations in comparing precedents also Berman
and Hafner (1995). In conclusion, ideally formal and computational models of
case-based reasoning should allow for any possible criterion for choosing between
conflicting precedents.
Nevertheless, similarity is certainly one of the most important criteria. In this
respect, the multi-step nature of judicial opinions introduces a complication not
treated by the HYPO model. If a precedent contains multiple issues, it is not the
similarity of the entire precedent to the new situation that must be determined, but
similarity of the portion of the precedent that pertains to the issue at hand (likewise
Branting, 1991, 1994).
2.2.3. A Dialectical Perspective on the Evolution of Case Law

In the debate on precedent, formalistic (strict) and anti-formalistic (sceptic) ap-
proaches are frequently opposed (cf. e.g. MacCormick, 1987, p. 157; Twining and
Miers, 1991, p. 311). The first approach construes the binding meaning of the
precedent on the basis of the text of the opinion and the plausible intention of
the judge. The latter approach looks beyond the text and its author, by considering
interpretations given by subsequent judges, and more generally, by providing a
holistic interpretation of the development of case law.
Our argumentation-based approach allows us to find a middle way between
these two perspectives. Each rule, being dialectically supplemented or limited by
rules contained in other cases, gains a certain degree of flexibility, without being
attributed an indeterminate content. We argue that frequently when the phrase 'the
rule of a case' is used, it does not stand only for the original rule deciding (an
issue of) that case but rather for a rule set including not only the original rule,
but also the analogies drawn for it, and the exceptions defeating it, in subsequent
cases. In this perspective, we can understand how subsequent decisions can modify
the scope of a precedent's ruling, although leaving the formulation of the original
rule unchanged. In particular, this corresponds to our understanding of the process
of restrictive distinguishing. As we have seen above, we do not need to model
this process by replacing the precedent's rule with a new, more restricted rule,
the antecedent of which also includes the complement of the factor justifying an
opposite outcome in the new case (Raz, 1989). This is since an equivalent result
is given in our framework by the dialectical interaction of the old rule and its new
exception.
So, when Donoghue was analogized into a rule establishing the responsibility
of repairers (or when the analogy concerning legal professionals was rejected) the
set of the rules directly or indirectly concerning Donoghue was extended. This
[ 141 ]
allowed (or blocked) new inferences in subsequent cases (the extended rule set
supports the responsibility of repairers and the non-responsibility of lawyers). This
set (or a subset of it) is dynamically constructed by subsequent citations, when the
new case satisfies the conditions of more than one Donoghue-related rule.
In this way our model is consistent with 'formalist' approaches such as (Mac-
Cormick, 1987, p. 170) or (Cross and Harris, 1991, p. 72) who point to those
rules or rulings being "expressly or impliedly given by a judge". At the same time
our model is also consistent with those theories where the 'meaning' of a case
changes in time, being "the rule of law for which a case is made to stand or is
cited as authority by a subsequent interpreter" (Twining and Miers, 1991, p. 312)
or Dworkin's view of opinions as chapters of a chain novel, to be continuously
reinterpreted (Dworkin, 1985, pp. 158 ff.).
2.3. SUMMARY
In sum, it appears that a complete formal or computational account of precedent-

based judicial reasoning should at least satisfy the following criteria. With respect
to individual precedents, such a model should be able to represent their dialectical
structure, where the judge considers arguments for and against the decision of an
issue. The model should also be able to represent the stepwise nature of precedents,
where the final decision is reached after resolving a series of intermediate issues.
Finally, the model should be able to represent multi-level arguments, including
arguments on the choice between conflicting arguments.
With respect to the use of precedents in solving a new case, the model should,
firstly, be able to capture the analogical use of precedents in a dialectical setting:
similar precedents can be cited, but those citations can be countered on the grounds
that the similarity is not sufficient, or that more similar or otherwise preferable
precedents point to the opposite outcome. Moreover, the model should be able to
determine the similarity of portions of precedents pertaining to one issue rather
than of precedents as a whole. Finally, it should leave room for other standards
besides similarity for choosing between conflicting precedents.
If these criteria are satisfied and the dialectical nature of cases and case-based
reasoning is captured, a middle way becomes possible between the formalistic
and anti-formalistic views on the ratio decidendi of a case, where the 'meaning'
of a case changes in time, determined by the analogies drawn to a case, and the
exceptions made to it.
3. Some Approaches to Case-based Reasoning in AI & Law

As already indicated in the introduction, one of the aims of this article is to build a
bridge between AI & Law research on case-based reasoning and more recent logic-
oriented research on defeasible argumentation. In particular, we shall consider
HYPO-style case-based reasoning, where a case is represented as a set of factors
[142 ]
pushing the case towards (pro) or against (con) a certain decision, plus a decision
which resolves the conflict between the competing factors; new cases are expected
to be resolved in accordance with decisions performed in the past. Our choice for
the HYPO approach is motivated both by its prominent place in AI & law research
on case-based reasoning and by its inclination toward dialectical argumentation (it
produces an alternating sequence of arguments of a proponent and an opponent of
a claim, where each argument attacks the previous argument of the other party).
We then discuss some systems which extend or modify HYPO's model, such as
Branting's model of ratio decidendi, CABARET and CATO. However, in order
not to drown the reader in too many details, we shall mainly focus on HYPO-
style analogizing and distinguishing. Nevertheless, it will tum out that some other
features of the above-mentioned systems can also be captured by our model.
3.l. REPRESENTING AND REASONING WITH CASES IN HYPO

HYPO aims to model how lawyers make use of past decisions in disputes with their
opponents. The system generates disputes between a proponent ('plaintiff') and an
opponent (,defendant') of a legal claim, where each move conforms to certain rules
for analogizing and distinguishing precedents. These rules determine for each side
which are the best cases to cite initially, or in response to the opponent's move, and
how the opponent's cases can be distinguished. A best case for a side is a case that:
has the disposition (decision) wished by that side;

shares with the Current Fact Situation (CFS) at least one factor which supports
that disposition;
shares a most inclusive set of factors with the CFS, in comparison with other
cases confirming the desired decision (on pointness).
A citation can be countered by a counterexample, that is, a case that is at least as

much on point, but has the opposite outcome. A citation may also be countered by
distinguishing, that is, by indicating a factor in the CFS which is absent in the cited
precedent and which supports the opposite outcome, or a factor in the precedent
which is missing in the CFS, and which supports the outcome of the precedent.
In choosing an appropriate argument move, a crucial aspect is the tendency
of facts towards or against a decision. Consider the following example, where the
issue is whether a stay in another country changes one's fiscal domicile with respect
to income tax. Assume that the following factors pro and con can be identified.
Pro change is that the old house was given up, while con change is that it was
kept.
Pro change is that the tax payer's company is based in the new country, while
con change is that the company is based in the old country.
[143 ]
Pro change is that the duration of the stay is long, while con change is that the
duration is short. 5
Now assume we have the following three precedents, where the factors that are pro
the decision 'change' are given in typewriter and the factors con the decision
'change' in italics (note that we do not assume that each factor receives a definite
value in each case: the duration may be neither long nor short, so that it does not
push the decision in any direction).
Prec A: Factors: long duration, gave up house,
domestic company
Decision: change
Prec B: Factors: foreign company,

kept house
Decision: change
Prec C: Factors: gave up house,

short duration, domestic company
Decision: no change
Assume that the facts of a new case (the CFS) are:
CFS: long duration, domestic company, kept house
All precedents share some factors with the new case:
Prec A n CFS = { long duration, domestic company}
Prec B n CFS = { kept house }
Prec C n CFS = { domestic company}
Suppose that Side 1 in the new case wants to argue that in the CFS the fiscal
domicile has changed. Although both Prec A and Prec B have this outcome, B
is not citable for Side 1, since the only factor it shares with the common situation
is against change; only A can be cited by Side 1. Side 2 can only answer to the
citation of A by distinguishing, that is, by referring to the factor kept house, which
is a con-change factor in the current situation not shared by Prec A, or by referring
to gave up house, which is a pro-change factor in the precedent not shared by the
CFS.
In evaluating the relative force of the moves, HYPO uses the set inclusion or-
dering on the factors that the precedents share with the CFS. For instance, if in the
above CFS a party cites Prec C in defence of the claim 'no change', then the other
5 For simplicity we assume in this paper, unlike HYPO, that all factors are two-valued, i.e., either
true or false.
[ 144]
party can distinguish C with respect to the factor 'long duration', by drawing an
analogy with the 'trumping counterexample' Prec A. The citation of A is regarded
as better ('more on point') than that of C, since the factors that A shares with the
CFS include those that C shares with the CFS.
Note that HYPO's reasoning forms (citing cases, citing counterexamples, dis-
tinguishing) are based not only on the set inclusion ordering on the shared factors
with the CFS, but also on the tendency of the factors toward a certain outcome. A
case is only citable for a side if this case shares with the CFS a factor favouring
that side, and a case missing a factor which is included in the CFS can only be
distinguished if the missing factor is against the outcome of the distinguished case.
Therefore, in representing precedents it is essential that this tendency of factors
be somehow represented. HYPO does so by simply marking them Pro or Con the
decision. In this paper we shall propose an alternative method.
In conclusion, HYPO's model is quite attractive, since it emphasizes the dialec-
tical nature of legal reasoning while reducing it to a limited set of argument moves,
based on a simple knowledge representation scheme. HYPO also addresses the
requirement that judicial rationales must have a dialectical multi-argument struc-
ture: this combination of conflicting arguments is implicit in the representation
of a case as a set of conflicting factors. Furthermore, HYPO implements ways of
analogizing and distinguishing and, finally, HYPO provides a way of comparing
conflicting precedents, by using the more-on-point ordering. Those are the aspects
of HYPO which we basically want to transfer to our model.
In some regards, however, we also want to extend and generalize HYPO's
model. In particular, it has frequently been observed that HYPO has no way of
representing how facts contribute to a decision. Cases are essentially represented as
a collection of factors, and a decision; no intermediate reasoning steps from factors
to decision can be represented. 6 Therefore HYPO does not do full justice to the
typical stepwise construction of legal arguments, discussed above in Section 2.1.2.
In capturing this feature of precedent-based reasoning, we shall adapt a proposal
of Branting (1994).
3.2. BRANTING'S MODEL OF RATIO DECIDENDI

Branting (1991; 1994) has proposed to represent the ratio decidendi of precedents
as a 'reduction graph', where more basic factors are linked by 'warrants' to more
abstract factors. From our logical point of view this boils down to representing a
case as a logical argument, i.e., as a logically valid sequence of reasoning steps
starting from a set of premises. In our example Branting would allow a precedent
to include a multi-step argument (concluding for the decision of the case), in which
the fact that the company has foreign headquarters (in the country where the em-
ployee is going to work) determines that it is a foreign company, which in turn
6 Although HYPO allows rules for determining the presence of a factor on the basis of factual
input, these rules cannot be used (or attacked) for dialectical purposes.
[145 ]
determines that the employee's fiscal residence is changed. Such an argument is

represented as a combination of rules. The top rule, or 'warrant', concludes to the
decision of the case (foreign-company A kept-house =} change). The lower level
ones are called reduction warrants: they reduce their consequent to the more basic
(more factual) conditions contained in their antecedent. For example, foreign head-
quarters =} foreign-company reduces foreign-company to foreign-headquarters.
Both the final warrant and the reduction warrants are citable, according to Brant-
ing, and, more generally, he admits the citation of parts ("portions") of the ratio
decidendi (as allowed in his GREBE system, Branting, 1991).
We want to borrow Branting's stepwise representation of precedents and his
idea to admit the citation of portion of precedents, and include them in our dialec-
tical model of HYPO-style reasoning. We do not address some other aspects of
Branting's model, since they are not directly relevant for our present purposes.
3.3. CABARET
The CABARET system of Skalak and Rissland (1991) has a different focus than
our model: it contains heuristics for combining statutory (or other) rules and prece-
dents in statutory interpretation, in particular for using precedents to confirm or
contest the application of a rule. An elaborate model of legal reasoning is cor-
respondingly provided. This model is further developed in (Skalak and Rissland,
1992), where argument strategies, moves and primitives are distinguished, and in
which the reasoner's point of view is, as in HYPO, essential. Although this work
is very interesting, it does not strictly adhere to a dialectical model of dispute of
the kind we want to study, and therefore we shall not go into a full description
of CABARET, but just comment on those features which directly address our
concerns.
Firstly, CABARET allows statutory rules to be analogized by broadening, in
particular by citing a case where some of the rule's preconditions were missing
but its conclusion was still upheld. And CABARET allows such analogies to be
countered by (non-restrictive) distinguishing. Furthermore, it allows for rules to
be contested (discredited), by citing a case where the opposite was decided (this
corresponds to what we called restrictive distinguishing). Finally, it allows cases
to be used for establishing the antecedent of a rule. Thus CABARET has a certain
mUlti-step structure, since a rule might have more than one antecedent. However,
each precedent is still represented as in HYPO, i.e., as a one-step decision.
Within our model we want to capture the just-mentioned reasoning forms. We
leave it to future research how other argument forms identified by Rissland and
Skalak can be modelled in our framework.
[146]
3.4. CATO
The CATO system of Aleven and Ashley (1996; 1997) is an intelligent learning
environment for teaching case-based argumentation skills to law students. Like
HYPO, CATO uses factors to represent cases. Its set of basic argument moves in-
cludes a number of HYPO's argument moves and contains additional ones as well.
Unlike HYPO, CATO is capable of organizing multi-case arguments by issues,
following a standard rhetorical format. A key element in CATO's architecture is a
so-called 'factor hierarchy'.
It is outside the scope of this paper to give a full account of the many interesting
aspects of CATO. Here we confine ourselves to the fact that it goes beyond HYPO
in addressing the stepwise construction of legal arguments. This is one of several
purposes for which CATO employs the factor hierarchy, which, like Branting's
reduction graphs, links more and less abstract factors. However, while Branting has
a different graph for each case, expressing the justification of the case's decision,
CATO has just one hierarchy, expressing expert knowledge about the domain. In
CATO's hierarchy lower factors are labelled according to whether they are a reason
pro or con the higher factors they are linked to. Links are labelled according to their
strength (weak or strong), which labels can be used to solve certain conflicts. Cases
are, as in HYPO, still represented as one-step decisions, but users of CATO can use
the factor hierarchy in several ways for interpreting the theory behind a decision.
For instance, it is possible to discuss the relevance of a distinction, i.e., to argue
why in the current situation the decision should be different, due to the distinction,
or why it should be the same, in spite of it. The corresponding argument moves are
called emphasizing and downplaying a distinction.
Emphasizing a distinction consists not only in pointing at the factual differ-
ences, but also in stressing that, according to the factor hierarchy, the additional or
missing distinctive factors allow the construction of a multi-step argument why the
distinction matters. For instance, assume a factor hierarchy for our tax example in
which (see Figure 1) two basic factors kept house and kept old car number plate
are both positively linked to the more abstract factor showed intention to return,
which in tum is negatively linked to change. And assume the following precedent.
Prec D: Factors: short duration, kept house
Decision: no change
Assume, furthermore, that the current fact situation is
CFS 1: short duration
HYPO allows to distinguish precedent D by pointing at the factor kept house of
Prec D, which is missing in the CFS. CATO allows in addition to emphasize this
distinction, by saying that therefore in the CFS no intention to return was shown,
unlike in precedent D. Thus the significance of the distinction is explained in terms
of the factor hierarchy. If, moreover, the CFS contains a factor con showed intention
[ 147]
change
I /19: showed-intention-to-retum
/~
/I: kept-house I /Is: kep t- old- car- number-plate
Figure 1. A partial factor hierarchy.
to return, CATO would have mentioned this as well, to draw an even stronger
contrast between the cases.
Downplaying a distinction exploits the hierarchy in a different way. It consists
in saying that in spite of the apparent distinction between the cases at the factual
level, at a more abstract level a parallel can be drawn. Assume that the facts are
now as follows.
CFS2: short duration, kept old car number plate
If now Prec D is distinguished since the CFS lacks D's factor kept house, then this
distinction can be downplayed by referring to the factor hierarchy, by saying that
both in Prec D and in the CFS there is evidence that showed intention to return.
(The reader will have recognized that downplaying a distinction in fact employs
the abstraction type of analogy, discussed above in Section 2.2.1.)
In our model we want to include the possibility of a factor hierarchy, but we
shall not try to directly model downplaying and emphasizing a distinction, since
the different aims of our model and of CATO (which is a tutoring system) seem
to prevent a straightforward inclusion of these moves in our approach. Instead we
shall briefly indicate how aspects of these moves, and of abstraction type analogy,
could be added in future research. In some respects we also want to extend and
generalize CATO's approach. In particular, while CATO's factor hierarchy is fixed,
we want to be able to assert multiple view points on the factor hierarchy and on the
priorities between factors, and to make those priorities dependent upon arguments.
[148 ]
4. The Building Blocks of our Proposal

In this section we present the main ingredients of our proposal. After a sketch
of the main ideas (4.1.), we outline some basic notions of our previously devel-
oped logical argumentation system (4.2.), introduce a new method for representing
precedents (4.3.), and present the dialectical form of our argumentation system
(4.4).
4.1. THE BASIC IDEA

As already indicated, we want to present, within the analysis of Section 2, a logic-
based model of dialectical case-based reasoning that preserves some strong points
of the above CBR systems, but that also extends and generalizes them in some
respects. Let us briefly summarize the basic requirements which our model intends
to satisfy.
As for the dialectical setting, we want to model the idea of normative di-
alectics: rules for dispute should reflect dialectical asymmetry between the
proponent and the opponent of a claim, and the aim of the dispute is to test
whether a claim is tenable.
As for the basic argument moves, they should contain broadening and restric-
tive and non-restrictive distinguishing as modelled in HYPO, they must allow
the use of portions of precedents (as in Branting's model), and they must allow
for broadening and discrediting legal rules (as in CABARET).
As for representing precedents, it should be possible to represent multi-step
arguments (as in Branting's proposal), dialectical structures, i.e., arguments
and counterarguments (as in HYPO and CATO), and multi-level arguments,
i.e., arguments on rule-priorities.
Finally, as for comparing arguments, we want to retain HYPO's more-on-
point ordering as one of the criteria, but leave room for any other criterion.
And the criteria must be debatable, just as any other legal claim.
We shall try to satisfy these requirements with the following three main ideas.
Firstly, the dialectical setting will be captured by the dialectical proof theory for
defeasible argumentation defined in (Prakken and Sartor, 1996b, 1997a). This proof
theory embodies the idea of normative dialectics: it has the form of a dialogue game
between a proponent and an opponent for a claim, and its output is whether the
claim is justified, defensible or overruled. However, its assumption that there is a
fixed pool of premises will be replaced by certain ways to introduce new premises,
which is our second main idea. In particular, we shall model two argument moves,
broadening and non-restrictive distinguishing, as heuristics for introducing new
premises into a dispute. It is important to note that our original proof theory now
applies to each stage in a dispute, evaluating it in a manner to be defined below
in Subsection 5.4. Our final idea is a new case representation method, adapting an
[ 149]
idea of Loui and Norman (1995): each precedent will be represented not just as
one logical argument but as a set of possibly conflicting arguments. Each of those
arguments may include multiple steps (as Branting suggests), but the conclusion
derived in each of those steps may be the matter of a dispute in which factors pro
and con that conclusion are resolved (as in HYPO). Moreover, our representation
shall include arguments on rule priorities, to express that (and sometimes also why)
certain factors were outweighed by other factors.
4.2. LOGICAL PRELIMINARIES

The logical background assumed in this paper is the argument-based system of
Prakken and Sartor (l996a; 1996b; 1997a) (although other systems with similar
features will do as well). In this section we briefly describe the basic elements
of the system, except its dialectical form, which will be discussed below in Sub-
section 4.4. The logical language is that of extended logic programming. i.e., it
has both negation as failure ('"") and classical, or strong negation (--.). We add to
this language one feature: each formula is preceded by a term, its name. Rules are
strict, represented with ---+, or else defeasible, represented with =}. The idea is that
strict rules are beyond debate; only defeasible rules can make an argument subject
to defeat. Accordingly, facts are represented as strict rules with empty antecedents
(e.g. ---+ gave-up-house). The input information of the system, i.e., the premises,
is a set of strict and defeasible rules, which we call an ordered theory (,ordered'
since, as explained below, we assume an ordering on the defeasible rules).
The following notions are all defined relative to a given ordered theory. Argu-
ments can be formed by chaining rules, ignoring weakly negated antecedents; each
head of a rule in the argument is a conclusion of the argument. Conflicts between
arguments are decided according to a binary relation of defeat among arguments,
which is partly induced by rule priorities. An important feature of our system is that
the information about these priorities is itself presented as premises in the logical
language, as in the foHowing two example rules.
r: rl is decided by higher court than r2 =} r2 -< rl
r': rl protects manufacturers /\ r2 protects consumers /\
consumer runs higher risks than manufacturer =} rl -< r2
Here x -< y means that y is preferred over x. Thus rule priorities are like any other
piece of legal information established by arguments, and may be debated as any
other legal issue.
Our relation of defeat is a weak notion: that Ar gl defeats Ar g2 does not exclude
that Ar g2 also defeats Ar gl; this can happen, for instance, when a conflict between
two incompatible rules is not resolved by the given rule priorities. If, however,
Argl defeats Arg2 but not vice versa, we say that Argl strictly defeats Ag2.
There are three ways in which an argument Arg2 can defeat an argument Argl.
The first is undercutting it, which occurs if a rule in Argl contains'"" L in its
[150]
body, while Arg2 has a conclusion L. For instance, the argument [rj: -+ p, r2:
p =?- q] (strictly) defeats the argument [r3: "'-' q =?- r] by undercutting it (note
that "-' L reads as 'there is no evidence that L'). The other two forms of defeat
are only possible if Argj does not undercut Arg2. One way is by excluding an
argument, which happens when Ar g2 concludes for some rule r in Ar gl that r is
not applicable (formalized as .....,appl (r)). For instance, the argument [rl: -+ p, r2:
p =?- .....,appl (r3)] (strictly) defeats the argument [r3: =?- r] by excluding it. The final
way in which Arg2 can defeat Argl is by rebutting it: this happens when Argl and
Arg2 contain rules that are in a head-to-head conflict and Arg2's rule is not worse
than the conflicting rule in Ar gl. For instance, the arguments 7
[fl: short-duration, rj,' short-duration =?- change-fiscal-domicile]
[12: foreign-company, r2:foreign-company =?- ....., change-fiscal-domicile]
defeat each other if the rules rl and r2 are of equal priority or if no priority relation
holds between them, while the first strictly defeats the second if rl has priority over
r2·
The defeat relation is not yet the final assessment of arguments: since it is just a
comparison between two individual arguments, it does not capture the phenomenon
of 'reinstatement'. To illustrate this notion, assume that an argument A is strictly
defeated by an argument B, which in turn is strictly defeated by an (undefeated)
argument C; then intuitively C reinstates A. This is captured by the final element
of our system, which, taking all interactions between the possible arguments into
account, divides these arguments into three classes: the justified arguments, those
with which a dispute can be 'won', the overruled arguments, with which a dispute
should be 'lost', and the defensible arguments, which should leave the dispute
undecided (recall that these notions are relative to a given ordered theory). The
proof-theoretical version of his definition has the form of a dialogue game, and
will be discussed below in Subsection 4.4.
4.3. A METHOD FOR REPRESENTING CASES

Our method for representing legal precedents has two components: representing the
tendency of a factor in a logical rule, and representing a precedent with conflicting
factors as a set of conflicting logical arguments. As for the tendency of factors, this
is represented as follows. A rule' f is a reason pro d' is represented as a rule
r:f=?-d
We may also have a conjunctive reason, as in the following rule.
r : fl /\ 12 =?- d
7 For notational convenience we shall often list a fact of a case as a literal preceded by a name Ii.
Formally, a fact Pi with name Ii is a strict rule Ii: -+ Pi.
[ 151 ]
As for representing precedents, above we said that we want to represent them as

collections of, possibly conflicting, arguments. In fact we shall for notational con-
venience present them in a slightly simpler way, viz. as sets of rules from which the
arguments pro and con can be constructed. The precise method is as follows. Since
we want to read a rule antecedent =9 consequent as saying that the antecedent is a
reason for the consequent, we cannot express a dominance of, say, the pro factors
over the con factors by conjoining pro and con reasons in the antecedent of a rule
with the pro consequent. Instead we use a representation which directly expresses
the tendency of each factor, and the resolution of their conflict. The simplest of
such formalizations consists in separately representing each reason statement, and
adding one or more rules concerning their comparative evaluation.
For example, let us assume that a short duration of the working stay outweighs
the fact that the company is foreign. We do not express that by a combined rule
r: short-duration /\ foreign-company =9 ...., change-fiscal-domicile
Instead, we represent the resolution of the conflict as a pair of conflicting rules,
together with a priority statement.
r] : short-duration =9 ...., change-fiscal-domicile
r2: foreign-company =9 change-fiscal-domicile
r3: antecedent =9 r2 -< rl
Here antecedent expresses the reasons why short-duration outweighs foreign-
company (as far as a change in fiscal domicile is concerned). In realistic examples
antecedent will itself often be derived (dialectically) from other rules. In fact,
one major advantage of this representation scheme is that it makes it possible
to express the grounds why certain factors override certain other factors, in the
form of antecedents of priority rules. See, for instance, the two priority rules in
Subsection 4.2, that had antecedents r] is decided by higher court than r2 and
rl protects manufacturers /\ r2 protects consumers /\ consumer runs higher risks
than manufacturer. As these examples show, the priorities can be based on any
ground, ranging from general legal principles to case-specific considerations. It is
not necessary (although possible) that they reflect certain general legal principles,
like 'higher courts precede lower courts' or 'later decisions prevail over earlier
ones'. And the priorities can very well depend on considerations that are specific
to the context of a case, as in the above rule antecedent about manufacturers and
consumers.
Finally, as usual in legal theory, we assume that precedents contain general
rules, i.e., that they contain variables instead of ground terms (except for the prior-
ity rule of a case).
[152 ]
4.4. THE DIALECTICAL CONTEXT
Just defining a representation method in a logical language is not enough; we must

also specify the dialectical context in which the represented information can be
used. As indicated in Section 4.1, we want to embed our representation method
in the dialectical proof theory that was developed in Prakken and Sartor (l996b,
1997a) for the just-explained system for defeasible argumentation, but we want to
replace its assumption that there is a fixed pool of premises (the ordered theory)
with ways of entering new information into a dispute. We now present the dialec-
tical proof theory; in the following section it will be embedded in a protocol for
premise introduction. For more technical details on the proof theory and related
work of others the reader is referred to (Prakken and Sartor, 1997a) and (Prakken,
1998).
The proof theory has the form of a dialogue game. Its purpose is to determine
whether a given formula defeasibly follows from a given ordered theory or not,
i.e., whether it is a justified conclusion on the basis of the ordered theory or not.
Thus the game is an example of normative dialectics, with the resulting dialectical
asymmetry between the players. A proof that a formula is justified takes the form
of a dialogue tree, where each branch of the tree is a dialogue, and the root of
the tree is an argument for the formula. Every move in a dialogue consists of an
argument based on some given ordered theory. Each stated argument attacks the
last move of the opponent in a way that meets the player's burden of proof. The
required force of a move depends on who states it. Since the proponent wants a
conclusion to be justified, a proponent's move has to be strictly defeating, while
since the opponent only wants to prevent the conclusion from being justified, an
opponent's move may be just defeating. A subtlety here is that for determining the
force of the proponent's move only the priorities stated by that move count, while
for assessing the opponent's move no priorities need to be applied at all.
Here is the central definition of the dialogue game (' Ar g-defeat' means defeat
on the basis of the priorities stated by Ar g).
DEFINITION 4.1 (dialogues) A dialogue is a finite nonempty sequence of moves

movei = (Playeri, Arg;) (i > 0), such that
1. Playeri = P iff i is odd; and Playeri = 0 iff i is even;
2. If Playeri = Playerj = P and i -=1= j, then Argi -=1= Argj;
3. If Playeri = P then Argi is a minimal (w.r.t. set inclusion) argument such
that
(a) Argi strictly Argi-defeats Argi-l; or
(b) Argi-l does not Argi-defeat A i - 2 ;
4. If Playeri = 0 then Argi 0-defeats Argi-l.
A dialogue is based on a set of rules r iff all rules of Ar gi are in r.
The first condition says that the proponent begins and then the players take turns,
while the second condition prevents the proponent from repeating a move. The last
[ 153 ]
two conditions form the heart of the definition: they state the burdens of proof for
P and O. Condition (3) gives P two types of moves: the first is an argument that
combines an attack on O's move with a priority argument that makes the attack
succeed; the second is a priority argument that neutralizes the defeating force of
O's last move. Finally, condition (4) says that 0 does not have to take priorities
into account.
The following simple dialogue illustrates this definition (the rule names refer
ahead to the example below in Section 6.4.).
PI: [fl: kept-hollse,
r]: kept-house:::} -, change]
0]: [flO: -, domestic-headquarters,

rIO: -, domestic-headquarters:::} -, domestic-company,
r4: -, domestic-company :::} change]
P2 : [h: domestic-property,
r7: domestic-property :::} domestic-company,
f16: r7 is decided by higher court than rIO,
p: r7 is decided by higher court than rIO :::} rIO -< r7]
The proponent starts the dialogue with an argument PI for -, change, after which
the opponent attacks this argument with an argument 0 1 for the opposite conclu-
sion. 0] has the required defeating force, since in our logical system two rebutting
arguments defeat each other if no priorities apply to the conflict. P2 illustrates the
first possible reply of the proponent to an opponent's move: it combines a 'nor-
mal' argument with a priority argument that makes it strictly defeat the opponent's
move. The second possibility, just stating a priority argument that neutralizes the
opponent's move, is illustrated by the following alternative move:
P~: [fl7: r] is more recent than r4,
pi: r] is more recent than r4 :::} r4 -< rI1
The point of this argument is to resolve the conflict between p] and 0] in favour
of Pl.
Next we recapitulate the definition of a 'dialogue tree'.
DEFINITION 4.2 (dialogue trees) A dialogue tree based on an ordered theory I

is a tree of moves such that
1. Each branch is a dialogue based on 1;
2. If Player; = P then the children of move; are all defeaters of Arg; based on
1.
[154 ]
won by P won by 0
Figure 2. Two dialogue trees.
The second condition of this definition makes dialogue trees candidates for being
proofs: it says that the tree should consider all possible ways in which 0 can
defeat a move of P. This is why the definition is relative to an ordered theory.
(Note that the definition further allows that P-nodes have several children). The
above example has (assuming there are not more premises) a dialogue tree of two
dialogues, PI - 0] - P2 and PI - 0] - P~.
The final definition summarizes when a player has won a dialogue and when an
argument and claim has been shown justified.
DEFINITION 4.3 (winning) A player wins a dialogue based on r iff the other
player cannot move. P wins a dialogue tree based on r iff he wins all its branches.
And 0 wins it iff he wins one of its branches. An argument A is justified on the
basis of r iff there exists a dialogue tree based on r with A as root and won by P.
And, based on r, an argument is overruled iff it is defeated by a justified argument,
and it is defensible iff it is neither justified nor overruled. Finally, a claim C is a
justified conclusion on the basis of r iff there is a justified argument for C based
on r.
Figure 2 illustrates the dialectical proof theory with two dialogue trees, where the
ordered theories, and the contents of the arguments are left implicit. The tree on
the right extends the tree on the left with one new branch, made possible by adding
new premises to the ordered theory. Arrows stand for defeat relations, so one-
directional arrows stand for strict defeat and bidirectional arrows for mutual defeat.
The dialectical asymmetry between P and 0 is reflected by the fact that all arrows
[155 ]
from P's moves to O's moves are one-directional, while some arrows from O's
moves to P's moves are bidirectionaL Assuming that the trees cannot be extended
with new arguments, the tree on the left is won by P, since all its branches end
with a move by P, so PI is a justified argument; by contrast, the tree on the right
is won by 0, since one of its branches ends with a move by 0, so here PI is not
justified.
This completes the overview of the dialogue game. In (Prakken and Sartor,
1997a) we prove that as a proof theory it is sound and under certain finiteness
conditions also complete with respect to the fixpoint semantics defined in (Prakken
and Sartor, 1996a). Since this semantics is a special case of the general framework
of Dung (1995) and Bondarenko et aL (1997), these results give our dialogue game
a well-understood logical foundation. In particular, they imply that the dialogue
game has some desirable properties. For present purposes the most important of
them are that the set of all justified conclusions is consistent, that the rule ordering
derived from these conclusions (and that also determines these conclusions) is a
strict partial order, and that the system satisfies the 'weakest link' principle that an
argument can only be justified if all its subarguments are also justified.
s. A Protocol for Reasoning with Precedents
The just-described dialogue game is defined relative to an arbitrary but fixed or-
dered theory. This is fine as long as the game serves as a nonmonotonic proof
theory, but when applied to the modelling of disputes, this is different, since in
actual disputes the premises are provided dynamically, in dialectical interaction
between the parties. However, in (Prakken and Sartor, 1996b) we remarked that the
dialectical proof theory also applies if the pool of premises is assumed to consist of
everything put forward by the parties in a dialogue. In the present section we want
to make this precise. More specifically, we want to regard HYPO-style reasoning
with precedents as heuristics for introducing information into a dispute, where each
premise introduction is contained in an argument that satisfies the player's burden
of proof as defined in Definition 4.1.
We shall discuss four things: the background information from which premises
can be constructed, reasoning with precedents, conducting actual disputes, and
evaluating them. It should be noted beforehand that the protocol is not meant to
be exhaustive: it does not want to capture all possible argument moves, but only
some of them, viz. HYPO-style reasoning with precedents.
5.1. THE BACKGROUND INFORMATION
We describe the protocol relative to a background theory of precedents, facts and

eventual other information. To this end we now first formally define the notion of
a precedent.
[ 156]
DEFINITION 5.1 (precedents) A precedent Case is a pair (CaseFacts, CaseRules),

where
CaseFacts is a set of strict rules;

CaseRules is a set of rules.
If Cases is a set of precedents, then Rules-oj-Cases is the union of the sets CaseRules
of all precedents in Cases.
Recall that in our model a precedent can be conceived in two ways, as consisting
of premises (facts and rules), or as consisting of arguments. The arguments in a
precedent Case = (CaseFacts, CaseRules) are exactly all those which can be built
from the ordered theory (CaseFacts, CaseRules). This means that those arguments
can with Definition 4.3 be classified as justified, defensible, or overruled on the
basis of r = Case. We believe that these notions may be relevant for the theory of
precedent, in particular for the distinction between rationes decidendi and obiter
dicta: the notion of a ratio decidendi of a case can be linked to that of (the rules
in) a justified argument, while that of an obiter dictum can be linked to that of (the
rules in) a non-justified one.
Next we define the 'background information' of the protocol. This should not
be confused with the ordered theory of our system for defeasible argumentation:
the background information is the information from which the ordered theory can
be (dialectically) constructed by the parties. It consists of a set of precedents, a
set of 'common-sense' rules, which might also include the applicable law (as in
CABARET) or a factor hierarchy (as in CATO), and a set of strict rules, represent-
ing the current fact situation.
DEFINITION 5.2 (Background Information) A Background InJormation theory

(B/) is a triple (Cases,CFS,CSRules), where
Cases is a set of precedents;

CFS is a set of strict rules, the current fact situation;
CSRuies is a set of rules, the 'common sense' knowledge.
5.2. REASONING WITH PRECEDENTS

How can the background information be used to introduce information into a de-
bate? As a first approximation the idea is that each move of the players should only
consist of rules from Rules-oj-Cases, CFS and/or CSRuies (obviously, the facts of a
precedent can in a new fact situation not be used). However, to capture HYPO-style
reasoning, we must also allow for the introduction of rules that are not contained
in any of these three sets but that can be obtained by analogizing or distinguishing
a precedent.
[157]
First we define how these kinds of reasoning can result in new rules. Informally,
the idea is that given a pre-existing (defeasible) rule we can produce two types of
rules: broadening rules, which have the same consequent as a pre-existing rule
but lack one or more of its antecedents, and distinction rules, which state that the
omitted antecedents are necessary for warranting that consequent. 8
We also use the following notation: for any rule r the set of literals occurring
in its antecedent is denoted by An t Li t s (r), while its antecedent and consequent
are denoted by ANT(r) and CONS(r). And for any set R = {rl, ... , rn} of
rules, AntLits(R) = AntLits(rl) U ... U AntLits(rn); likewise for ANT(R)
and CONS(R).
We now define how to broaden a rule. The idea is simple: a rule can be broad-
ened by deleting one or more of the literals in its antecedent.
DEFINITION 5.3 (broadening a rule). A defeasible rule r broadens a defeasible

rule r' iff
1. the first argument of r's name is r'; and
2. rand r' have the same consequent; and
3. AntLits(r) ~ AntLits(r').
In our model we allow two forms of distinguishing, which can be called weak and
strong distinguishing. 9 Weak distinguishing just concludes to the inapplicability
of a broadening rule, i.e., it excludes the argument using that rule, while strong
distinguishing argues that the opposite conclusion holds, i.e., it rebuts the argu-
ment using the broadened rule. The precise definition is technically more involved
than the one of broadening, but the basic idea is simple. If the other party has
broadened a rule by omitting one or more literals in its antecedent, then one can
strongly distinguish by saying that if the omitted literals cannot be proven, the
opposite conclusion holds, and one can weakly distinguish by saying that then the
broadening rule is inapplicable. This is formalized by giving the following content
8 To formalize this, we must also define a naming convention for rules, in order to encode which
rule broadens which. As for the rule names of the case rules, we assume that these consist of function
expressions of the form
name(t, tl, ... , tn )
where the function symbol name is as usual the informal name of the rule, and where the
terms tl, ... , tn are as usual the terms occurring in the rule. What t stands for depends on whether
the rule is intended to be a broadening of another rule or not. If not, then t = name; otherwise t is
equal to the name of the rule of which name is intended to be a broadening. This condition is needed
to deal with situations where a certain rule r contained in one precedent CI coincides with a rule in
another precedent C2. If r is directly used as contained in CI, the first argument of r's name is r,
and if r is used as broadening a rule r' in C2, the first argument of its name is r'. Only in the second
case is r distinguishable. Finally, we assume that rules that are the same except differences in their
terms have the same informal name. In our examples we leave the convention implicit and only give
the function symbol part of the names (e.g. rt instead of rt (name, x, y».
9 In (Prakken and Sartor. 1997b) we only formalized strong distinguishing.
[ 158 ]
to a distinction rule d concerning a broadening b. The antecedent of d contains

the weak negations of the literals that were omitted in b, and the consequent of d
is either the complement of b's consequent (strong distinguishing) or of the form
-.appl(r) (weak distinguishing).
DEFINITION 5.4 (distinguishing a rule). A defeasible rule r strongly distinguishes

a defeasible rule r' iff
1. rand r' have contradictory consequents; and
2. there exists a defeasible rule r" broadened by r' and there exist literals L 1, ... Ln
such that
(a) L 1 , ••• Ln are included in ANT(r") but not in ANT(r');
(b) ANT(r) = rv Ll /\ ... /\ rv Ln(n > 0).
A rule r weakly distinguishes a rule r' iff
1. The consequent of r is -.appl (r'); and
2. Condition (2) of strong distinguishing holds.
Note that the antecedent of the distinguishing rule is a conjunction of weak literals
r v L 1 /\ ..• /\ rv Ln: thus the distinction says that if L 1 /\ •.. /\ Ln are not proved,
then the consequent of the attacked broadening (and of the original rule) does not
hold. The literals Ll /\ ... /\ Ln are exactly those literals which were cancelled
from the original rule r" in order to produce the broadening r'. In other words,
when distinguishing, one claims that without the missing conditions (required by
the original rules and omitted in the broadening) the conclusion of the broadened
rule does not hold.
To give a simple example, assume that a case has a rule lO
r": a /\ b ~ c
and assume that r" is broadened by deleting b from r"'s antecedent, resulting in
r': a ~ c
Then r' can be strongly distinguished by weakly negating the missing literal b and
adding it to the antecedent of a rule with a consequent opposite to that of r':
r: rv b ~ -. c
Correspondingly, r' can be weakly distinguished by a rule with the same an-
tecedent but with a consequent stating that r' is inapplicable.
r: rv b ~ -. appl(r')
Here is a final example. Consider the rule
r2j4j6: -. kept-house /\ -. domestic-company /\ long-duration ~ change
The following rule broadens r2j4j6:
lOIn the rest of this paper we leave the function arguments of the rule names implicit.
[ 159]
r4/6: -, domestic-company /\ long-duration =} change

while the following rule (strongly) distinguishes r4/6:
rd2: '" -, kept-house =} -, change
Note that our definition of distinguishing only covers non-restrictive distinguish-
ing, i.e., cases where the CFS misses some factors of the precedent; restrictive
distinguishing, i.e., emphasizing a new factor not present in the precedent, can be
modelled by adding knowledge about the tendency of factors to CSRules: for each
Jactor for supported conclusion we can add a rule
r: Jactor =} supported conclusion
to CSRules. Any such rule can, when its antecedent is satisfied, be cited to counter a
conflicting precedent rule. Finally, HYPO's notion of a counterexample, i.e., citing
a case with the missing factor and the opposite outcome, is captured in our model
by the possibility of simply using the counterexample as a counterargument, as will
be illustrated below in Section 6.4.
5.3. ACTUAL DISPUTES

As stated several times above, our dialogue game of Section 4.4. serves as a (dialec-
tical) proof theory, and therefore it assumes a fixed set of premises. In the present
section, however, we are concerned with actual disputes, i.e., disputes in which
the set of premises is constructed dynamically, during the dispute. Accordingly, we
now define the notion of an actual dialogue, i.e., a dialogue as it can actually evolve
between the parties in a dispute. This boils down to defining the precise content of
the ordered theory r referred to in Definition 4.1.
We first define some useful notation. Let us denote for any BI = (Cases, CFS,
CSRules) the set of all broadenings of any rule in Rules-oj-Cases with Broaden-
ingsBl (note that Rules-oj-Cases U CSRuies S; BroadeningsBl) and the set of all
rules distinguishing any rule in BroadeningsBI with DistinctionsBI. Then we define
the introducible rules (on the basis of BI) as follows.
DEFINITION 5.5 (introducible rules).
IntroduciblesBl ;2 BroadeningsBI U DistinctionsBI.
Thus the set of introducible rules does not only contain all rules of any case in
CasesBI but also all rules that can be formed by broadening any ofthose rules, and
by distinguishing any of those broadenings. Note that this definition does not say
that the set of introducible rules is equal to the possible broadenings and distinc-
tions, but instead that these broadenings and distinctions are a subset of the set of
introducible rules. Together with the following definition this formalizes that our
protocol leaves room for other ways of introducing premises into a dispute.
[ 160]
We can now define an actual dialogue, and the corresponding notion of an

actual dialogue tree, as follows. The idea is that an actual dialogue conforms to
the definition of a dialogue game in Definition 4.1 with one important difference:
the rules in an argument do not have to be directly contained in a given set of rules;
they may also be constructed from these rules by broadening or distinguishing (or
they may be defined introducible in some other way).
DEFINITION 5.6 (actual dialogues) For any BI
- An actual dialogue based on BI is a dialogue D conforming to Defini-

tion 4.1 and based on CFS U CSRuies U Introducibles BI.
- An actual dialogue tree based on B I is a tree of actual dialogues based on
BI.
Note that we do not require that an actual dialogue tree contains all possible moves
of the opponent.
An actual dialogue tree should not be confused with a dispute as it actually
takes place. Rather, such a tree is a data structure which is built by the moves of
the parties in a dispute. Below we assume that a disputational protocol exists for
constructing an actual dialogue tree. However, since various such protocols are
conceivable (e.g. 'extend all nodes at once' or 'extend one node'), we shall not
define a particular one. We confine ourselves to observing that any such protocol
should allow for 'backtracking', i.e., for extending not only the leaves of the tree
but also earlier nodes. This is necessary since it may be that a player's move intro-
duces premises with which the other player can construct a new counterargument
against an earlier move of the first player (this cannot happen in the dialectical
proof theory of Section 4.4., where the ordered theory r is fixed).
5.4. WINNING A DISPUTE
Finally we define the outcome of a dispute. Assume that after movei in a dispute
an actual dialogue tree T; is constructed. Then the question is, who is winning at
stage i, if any? Our answer should maintain the link with Definition 4.3 and thus
with the semantics of (Prakken and Sartor, 1997a). Now Definition 4.3 is relative
to a given ordered theory, while the idea of our actual dialogues is that this ordered
theory is constructed dynamically. So we have to define what the content is of the
set r referred to in Definition 4.3.
Since in the present paper we focus on reasoning with precedents, we assume
for convenience that CFS and CSRuies are fixed, i.e., that r at least contains these
sets (although in other reasoning contexts they might also be constructed dynam-
ically). Then two alternative contents of r suggest themselves. The first contains
besides these sets all rules that are introducible: i.e.,
(1) r = CFS U CSRules U IntroduciblesBI
[ 161 ]
The second only contains the introducible rules that have actually been introduced
into the dispute. Accordingly, we now index r with the stage of the dispute (for
any tree T of moves, RulesT is the set of all rules occuring in T).
(2) ri = CFS U CSRuies U Rules-oj-Cases U (lntroducibles BI n RulesT;)
These definitions differ in the following way. The first formulates an ideal stan-
dard for disputes, requiring that the parties analogize and distinguish the available
precedents in the best possible ways. It does so because of the requirement in Defi-
nition 4.2(2) that a dialogue tree contains all possible moves of the opponent. When
combined with clause (1) just given, this means in particular that a dialogue tree
contains all possible distinctions that an opponent can make. The second definition,
by contrast, evaluates debates relative to the analogies and distinctions that have
actually been made, since it only includes in T; those broadenings and distinctions
that have actually been introduced by one of the parties upto T;.
In our opinion there is no need to choose which is the 'right' definition; both
definitions may have their uses, depending on the context.
Finally, we can define the notion of winning an actual dialogue tree (leaving the
content of r ambiguous between (1) and (2)). The definition directly states the link
with the proof theory of Definition 4.3 and thus indirectly with the semantics of
Prakken and Sartor (1997).
DEFINITION 5.7 (actually winning) For any actual dialogue tree T;:
l. P wins T; ifthere is a dialogue tree on the basis of r i with the same root as T;,
won by P, and containing only arguments of T;;
2. 0 wins T; if there is no dialogue tree on the basis of r i with the same root as
T;, won by P;
3. otherwise, T; is undetermined.
It is desirable to restate this definition in terms of the structure of T;, so that for
checking the result no new dialectical proof tree has to be constructed. It is easy to
see that P wins T; iff, by breaking off only branches after a move by 0 (so at a
choice point for P), T, can be pruned into a dialogue tree T;' of which all leaves are
P-moves and that cannot be extended after any P-move without introducing new
premises. For a win by 0 similar conditions on the structure of T; can be given.
Observe also that if r; is defined as in (1) above, the words 'without introducing
new premises' can be omitted.
It is important to note that when the proponent uses a broadening rule, this rule
can always be distinguished by the opponent. Therefore, if the opponent makes
optimal use of its resources, a proponent can only win a debate if it does not
draw analogies. This seems realistic, since with analogies the decision whether
to regard the similarities or the differences as more important is in the end a matter
of substance rather than of logical form.
At first sight, this observation would seem to reveal a drawback of our model,
since it would seem to imply that whenever the parties disagree, we cannot say
[162]
that a precedent 'controls a case'. Yet this is not true: the key is to take the third,
procedural level of legal argument into account. Our present model addresses the
fourth layer, i.e., it studies heuristics for premise introduction. However, in legal
reality disputes take place in the context of a legal procedure, and such a procedure
defines, among other things, how an actual dispute terminates, and how an arbiter
(the judge) has the procedural power to evaluate the introduced arguments, for
instance to decide whether the similarities or the differences between two cases are
more important. Now the crucial observation is that such an evaluation could be
modelled as a (final) premise introduction, viz. as one or more priority arguments,
after which Definition ?? can be used once more to compute the final outcome. In
sum, a full model of reasoning with precedents should also address the procedural
aspects oflegal reasoning (as argued before by Hage et at., 1994 and Gordon, 1995
and also by Berman and Hafner, 1991).
6. Applying the Protocol

In this section we illustrate how our dialectical protocol can be applied to HYPO-
style reasoning with precedents. We first discuss the adaptation of HYPO's more-
on-point ordering to our model, then explain some subtleties concerning distin-
guishing and combining factors, and then present an example. Readers wishing to
avoid too many technicalities can immediately proceed to Subsection 6.4.
6.1. ON-POINTNESS
In our model we want to use HYPO's similarity ordering on cases (the 'more on
point' ordering) as one of the sources of rule priorities (note that in HYPO it
is used for other things, such as selecting the best case to cite). In particular, if
two counterarguments cite rules from different cases, we want to give priority to
the rule from the case that is more similar to the current situation. However, it is
not obvious whether we can directly apply HYPO's more-on-point ordering, since
this ordering was not defined for multi-steps precedents. In (Prakken and Sartor,
1997b) we promised that the present article would redefine HYPO's definition.
However, the problem has turned out to be more complex than we then realized,
and therefore we now confine ourselves to identifying the problems and briefly
indicating possible solutions.
Recall that HYPO's definition of the more-on-point ordering has two elements.
Firstly, it defines the on-pointness of a precedent as being the overlap between the
facts of the precedent and the current fact situation; and secondly, it determines
which precedent is more on point by ordering the various overlaps in terms of set
inclusion: precedent A is more on point than precedent B, if A's overlap with the
current situation is a superset of B's overlap. As for the second element, we shall
model it in the same way as in HYPO, i.e. according to set inclusion. However,
as for the first element, determining the overlap between the precedent and the
[ 163]
current situation, it is not obvious that we can, as in HYPO, simply intersect all
'input' facts of the precedent with the CFS. The reason is that our multi-steps
representation allows the citation of portions of precedents, i.e., the citation of a
precedent for just one of its decisions. This has two important consequences for
determining on-pointness. When a case is decided with a multi-steps argument:
the court often considers an input fact in only some of these steps; so whether
a case fact is relevant depends on the decision for which the case is cited;
in 'later' steps the court often considers not just facts but also the conclusions
of 'earlier' steps, acting as 'facts' for a later decision; so on-pointness often
depends not only on the facts of a case, but also on intermediate conclusions
of a precedent.
Let us illustrate this with an example. Assume we have two conflicting precedents
A and B. (The rule names refer ahead to the example in Subsection 6.4.)
A= {r7: domestic-property => domestic-company,
rlO/12: -. domestic-headquarters 1\ -. domestic-president
=> -. domestic-company,
r4/6: -. domestic-company 1\ long-duration => change,
rl: kept-house => -. change,
PI: => r7
-< rlO/12,
P2: => rl
-< r4/6,
fl: kept-house, f6: long-duration,
h: domestic-property, flO: -. domestic-headquarters,
f12: -. domestic-president}
A has an intermediate decision -. domestic-company and a final decision change.
B= {r7/9: domestic-property 1\ domestic-headquarters =>
domestic-company,
r3: domestic-company => -. change,
r2: -. kept-house => change,
PI: => r2 -< r3,
12: -. kept-house, 19: domestic-headquarters, h: domestic-property}
B has an intermediate decision domestic-company and a final decision -. change.
Consider, furthermore, the following current fact situation.
CFS = {12: -. kept-house, h: domestic-property,
flO : -. domestic-headquarters}
Assume that A is cited by the proponent in a dispute as follows, broadening rlO/12
into blO and r4/6 into b4.
[ 164]
Pi = {blO: ...., domestic-headquarters => ...., domestic-company,
b4 : ...., domestic-company => change
flO: ...., domestic-headquarters}
And assume that the opponent replies by citing B as follows, broadening r7/9 into
b7 .
OJ = {b7: => domestic-company,
domestic-property
r3: domestic-company => ...., change
h: domestic-property}
Which precedent is more on point? If, as in HYPO, we compare all input facts
in each cited precedent with the CFS, neither of the two precedents is more on
point than the other: precedent A shares with the CFS factors h and flO, while
precedent B shares 12 and h. However, this is a global comparison, not depending
on the decision for which a precedent is cited; what is also possible is comparing
the precedents relative to their decisions.
Let us make such a 'local' comparison of the precedents, and consider the
intermediate issue domestic-company. The aim of this comparison is to establish
a priority relation between b7 and b lO . As remarked above, we must identify the
facts that were considered by the court when it decided this issue. One possible
answer to this question is that these are the antecedents of the rules that have this
issue in their consequent. Now in precedent A these are the rules r7 and rlO/12.
The antecedents of these rules contain h, flO and !I2, of which h and flO still
hold in the CFS. Checking the same for precedent B, we see that its only rule
about domestic-company is r7/9, of which the antecedent contains h and 19. Of
these two facts, only h still holds in the CFS. The result is that in this approach
A is, with respect to the issue domestic-company, more on point than B, for which
reason blO has priority over b7.
In conclusion, restricting the similarity comparison to a portion of a precedent
can change the similarity assessment. In our example this is since the fact...., kept-
house, which in a global comparison prevents A from being more on point than
B, is in B not considered for the intermediate issue domestic-company but for the
final conclusion change.
Let us now tum to the problems in formalizing this analysis. One problem arises
when on-pointness is determined with respect to a 'later' conclusion in a precedent,
for instance, with respect to A's conclusion change. Suppose A is compared with
the following conflicting precedent.
C= {rl/18: kept-house 1\ kept-old-car-numberplate => ...., change,
r6: long-duration => change
PI: => r6 -< rl/18,
12: kept-house, !Ig: kept-old-car-numberplate, f6: long-duration}
[ 165 ]
As for A, we must now look at the antecedents of rl and r4/6, which are ....., domestic-
company, kept-house and long-duration. The point is that one of these literals, viz .
....., domestic-company, was derived in A from another rule, rlO/12, so we cannot
make a simple comparison between the facts of the case and the CFS. Instead, we
must check which of the literals can still be derived in the new situation.
Now the problem is to make the terms 'derived' and 'new situation' precise.
As for 'derived', does it mean that the literal must be the conclusion of just an
argument or of a justified argument? And as for the 'new situation', does it include,
besides the CFS, only the rules of the cited precedent or also the rules introduced at
a certain state of a dispute? We are inclined to answer both questions in the second
way, but the formalization is tricky.
Apart from this, there is another problem. It seems that even with respect to later
steps in a precedent it is sometimes better to look at the facts of the precedents (as
in HYPO and CATO) than at the antecedents of the rules about the later issue (as
we did above). Consider in addition to C the following precedent and current fact
situation.
D= {r4/6:""" domestic-company 1\ long-duration =? change
14: ....., domestic-company, 16: long-duration}
CFS: {f6: long-duration, 12: kept-house}

Since both C and D are one-steps decisions, our and HYPO's method give the
same outcome that C is more on point than D. However, assume that in another
case with the same facts as C, the judge has taken the same decision but has made
the underlying theory implicit:
E= {r[/Is: kept-house 1\ kept-old-car-numberplate =?
showed-intention-to-retum
r19: showed-intention-to-retum =? ....., change,
r2: long-duration =? change
P2: =? r2 -< r19,
12: kept-house, liS: kept-old-car-numberplate, 16: long-duration}
Then in the same CFS our method gives a different answer than HYPO, viz. that
E is not more on point than D. However, this seems less than obvious, since C
and E have precisely the same facts and the same outcome, and the CFS is also the
same; the only difference is that in the second case the judge has made the theory
underlying his decision explicit.
In conclusion, it seems that the question how to determine similarity between
cases cannot be answered with one single method. It might even be concluded
that there is room for disagreement as to how similarity should be measured. IO
Therefore, we assume in the rest of this article that the method for determining
10 This idea also seems to undedy CATO's 'downplaying a distinction' move.
[166]
on-pointness is chosen by the parties in a debate, and that the result is expressed in
the form of defeasible rules. Thus the similarity assessments become debatable.
Below we assume that the parties' on-pointness assessments are expressed in
the form
:::} (-.) More-on-point(Precl, rl, Prec2, r2)
which reads as 'with respect to the issue dealt with by rl of Preci and r2 of Prec2,
the first precedent is (not) more on point than the latter.' Note that this reading
makes the on-pointness relation relative to an issue treated by the precedents,
instead of to the precedents as a whole.
In the present context we assume that such on-pointness statements are con-
tained in CSRules. Since they are defeasible, they are open to attack, which captures
that the similarity criteria are debatable.
As remarked above, the idea is that similarity assessments induce a priority
relation between the rules that they mention (here r2 -< rl). This can be formalized
by ading the following rule to CSRules.
mop: More-on-point(Precl, rl, Prec2, r2) :::} r2 -< rl
Note that this rule is also defeasible, so that, even if the parties agree on the sim-
ilarity criteria, the more-on-point priorities can still be defeated by other priority
considerations.
6.2. DISTINGUISHING PORTIONS OF PRECEDENTS

The example in the previous subsection can be used to explain a subtlety with re-
spect to distinguishing. Argument Pi, which cites precedent A, uses two broadened
rules, so it is distinguishable in two ways. The choice how to distinguish is deter-
mined by which conclusion of Pi one wants to attack. If one wants to attack the
intermediate conclusion domestic-company, precedent A can be (weakly) distin-
guished with respect to factor 112: -. domestic-president, with a rule "-' -. domestic-
president:::} -.appl(b lO ). If instead one wants to attack the final conclusion change,
precedent A must be distinguished with respect to the intermediate conclusion -.
domestic-company, with a rule "-' -. domestic-company:::} -.appl(b4 ).
6.3. INTERACTING FACTORS

HYPO embodies an independence assumption with respect to factors in the sense
that adding factors pro and deleting factors con a decision always makes an ar-
gument for the decision stronger. However, in general this assumption is not war-
ranted. Firstly, it is not always the case that the combination of several factors with
the same tendency also has that tendency. To reuse an example of (Prakken and
Sartor, 1996b), even if rain and heat are individually reasons not to go jogging,
then the combination of these two factors might very well be instead a reason to
[ 167]
go jogging. Moreover, even if the combination of two factors does preserve their
tendency, it might do so with weaker force. In our example, even if the combination
of rain and heat is still a reason not to go jogging, it might be a weaker reason than
just rain or just heat, because the combination is less unpleasant.
The argument-based system that underlies our model respects these observa-
tions. Firstly, factors pro do not automatically combine into a new factor pro: two
rules fl =} d and h =} d do not logically imply a third rule fl 1\ h =} d; if
It 1\ h is also a reason for d, the third rule must be added by hand to the premises.
Moreover, even if this is added, its priority relations do not logically depend on
those for the individual rules. Thus, the logic of our system does not compel us to
recognise that by joining all reasons for a conclusion we always obtain a stronger
argument.
However, in those contexts when this is considered appropriate, this result can
be obtained 'by default', viz. by adding the following scheme of general rules to
CSRules. In this scheme r+ denotes any rule obtained from r by adding zero or
more literals to r's antecedent - note that by convention these literals are also
factors pro the conclusion - and r- denotes any broadening of r.
r:
Note that this rule scheme is defeasible, so that it can be defeated when appropriate.
6.4. AN EXAMPLE
We now give an example illustrating our theory, with a background information
theory based on our tax example.
THE BACKGROUND INFORMATION
We assume that the B I contains the following factors, which are added to CSRules
as rules factor =} supported conclusion.
[ 168 ]
Factor: supported conclusion:
fl: kept-house -. change

12: -. kept-house change
h: domestic-company -. change
f4: -. domestic-company change
fs: short-duration -. change
f6: long-duration change
h: domestic-property domestic-company
fs: -. domestic-property -. domestic-company
19: domestic-headquarters domestic-company
flO: -. domestic-headquarters -. domestic-company
fll: domestic-president domestic-company
f12: -. domestic-president -. domestic-company
f13: -. domestic-job-prospects change
f14: domestic-citizenship -. change
!Is: -. domestic-citizenship change
We also assume that CSRuies contains the mop rule of Section 6.1, the relevant
unconditional on-pointness rules, and the r rule of Section 6.3.
THE CASE BASE
The precedential knowledge base Cases consists of three cases. As for notation,
the rand p rules are in CaseRules and the f rules are in CaseFacts.
The first precedent is for -.change.
A = {r7: domestic-property => domestic-company,
rlO/12: -. domestic-headquarters 1\ -. domestic president =>
-. domestic company,
r3/S/14: domestic-company 1\ short-duration 1\
domestic-citizenship::::} -. change,
r2/13: -. kept-house 1\ -. domestic-job-prospects => change,
PI: => r2/13 -< r3/S/14,
P2: => rlO/12 -< r7,
12: -. kept-house, h: domestic-property, fs: short-duration,
flO: -. domestic headquarters, f12: -. domestic president,
f13: -. domestic-job-prospects, f14: domestic-citizenship}
[ 169]
This precedent includes ajustified argument for -'change, viz. [f5, 17, f14, r7, r3/5/14].
The second precedent is for change.
B = {r2: -, kept-house =:} change,
r5: short-duration =:} -, change,
P3: =:} r5 -< r2,
12: -, kept-house, f5: short-duration}
This precedent contains a justified argument for change, viz. [12, r2].
The third precedent concerns a company tax case, and includes a justified argu-
ment for -, domestic-company.
C = {rIO: -, domestic-headquarters =:} -, domestic-company,
r7 /11: domestic-property 1\ domestic-president =:}
domestic-company,
r~: -, domestic-company =:} -, taxable-company,
P4: =:} r7/11 -< rIO,
17: domestic-property, flO: -, domestic-headquarters,
fll: domestic-president}
Finally, the current fact situation is
CFS = {h: -, kept-house, 17: domestic-property,

f5: short-duration, flO: -, domestic-headquarters,
f13: -, domestic-job-prospects, !J4: -, domestic-citizenship}
A DISPUTE
We now discuss an example dispute based on the just-given case base and CFS.
The outcome of the dispute is evaluated according to the second definition of r
given in Section 5.4, i.e. relative to the actually introduced premises. The dispute
is graphically displayed in Figure 3.
The dispute starts with the ordered theory ro = CFS U Rules-of-Cases U CSRules.
The proponent (the tax office) wants to defend -, change and starts the dispute by
referring to A. In particular, the tax office wants to use the rule
r3/5/14: domestic-company 1\ short-duration 1\
domestic-citizenship =:} -, change

However, this rule cannot be directly cited, since the condition domestic-citizenship
is not satisfied in the CFS. An analogy is required, through which the broadening:
r3/5: domestic-company 1\ short-duration =:} -, change
[ 170]
analogizing A
PI:
[r3/5: domestic-company 1\ short-duration => -, change,
r7: domestic-property => domestic-company,
17: domestic-property, /5: short-duration]
counterexample to counterexample to
A by analogizing A by analogizing C
B (wrt change) (wrt domestic com-
any)
distingl ishing
A (wrt hange)
0 1: 0;:
[/2: -, kept-house, [rIO: -, domestic-headquarters =>
r2: -, kept-house => change] -, domestic-company,
/10: -,
domestic-headquarters]
a"·
I'
[rd14 : ~ domestic-citizenship => -, appl(r3/5)]
comparing comparing PI
on-pointness and 0; by citing
of PI and 01 A '5 priority rule
P2:
P~:
[ml: More-on-point(A, r3/5, B, r2),
mop: More-on-point(A, r3/5, B, r2) => (P2: => rl0/12 -( r7,
r2 -( r3/5] r: rl0/12 -( r7 => rIO -< r7j
Figure 3. A dispute (won by 0).
is inputted into the dispute (using Definition 5.3). Furthermore, P must also use
rule r7 of A, which provides the intermediate conclusion domestic-company. The
whole argument is
PI: [r3/S: domestic-company /\ short-duration:::} ...., change,
r7: domestic-property :::} domestic-company,
h: domestic-property, Is: short-duration]
At this point the ordered theory is II = 10 U {r3/S}.
In order to build an actual dialogue tree according to Definition 5.6, the oppo-
nent must now by clause (4) of Definition 4.1 state an argument that defeats PI. One
way in which 0 can do so is by giving a counterexample to the precedent A used
by PI, by referring to precedent B (this move was in Section 2 called 'restrictive
[ 171 ]
distinguishing'). No broadening is required since B contains a rule which directly

applies in the CFS.
01: [12: -, kept-house, r2: -, kept-house =} change]
According to our framework, 01 defeats PI, since its rule r2 head-to-head conflicts
with PI'S rule r3/5 (recall that for assessing the force of a's moves no priorities are
needed).
As for the ordered theory, we now have r 2 = r I, and the actual dialogue tree
constructed at this stage is T2 = PI - 01. Although this tree ends with a move
by 0, P can reply without introducing a new rule, so Definition ?? tells us that at
this point there is no winner. In particular, P can reply by stating a more-on-point
assessment (which we assume to be in CSRules) and by using the mop rule (which
is also in CSRules).
P2: [ml: More-on-point(A, r3j5, B, r2),
mop: More-on-point(A, r3/5, B, r2) =} r2 -< r3/5]
P2 is an application of Clause (3b) of Definition 4.1: it is a priority argument that
makes PI strictly defeat 01.
P's on-pointness assessment follows the local comparison method explained
above at the beginning of Subsection 6.1. According to this method, the court in A
has, when deciding -, change, considered the antecedents of r3/5/14 and r2/13, which
are -, kept-house, short-duration, -, domestic-job-prospects, domestic-company
and domestic-citizenship. Of these literals, the first three are as facts in the CFS,
while domestic-company can be derived from r7 with 17, which is also in the CFS.
Only domestic-citizenship does not hold in the new situation, so the 'overlap' of A
with the new situation is
{-, kept-house, short-duration, -, domestic-job-prospects,
domestic-company}
Applying the same method to B means that when deciding change, the court has
considered the antecedents of r2 and r5, which are -, kept-house and short-duration.
Both of these literals are in the CFS, so the overlap of B with the new situation is
{-, kept-house, short-duration}
Clearly, this set is a strict subset of the overlap of A with the new situation, so with
respect to change A is more on point than B.
Note that in HYPO the fact that 01 is less on point than PI makes that 01 is not
an allowed reply to PI. The reason why we instead allow this move is that perhaps
a can attack P with a conflicting priority argument, based on grounds other than
similarity (although in our example a cannot construct such an argument).
We now have that r3 = r 2, and the actual dialogue tree at this point is T2 =
PI - 01 - P2 . Although its only branch ends with a move by P, P is not winning,
since a can extend the tree without introducing a new rule: a can create a new
[ 172]
branch of the tree by citing a portion of case C, which contains an argument that
attacks a subargument of PI.
Oi: [rIO:"'" domestic-headquarters::::} ...., domestic-company,
110: ...., domestic-headquarters]
We now have that r 4 = r 3 , and the actual dialogue tree at this point has two
branches: T3a = PI - 0 1 - P2, and T3b = PI - Oi. Nobody is winning, since the
proponent can again extend the tree without introducing a new rule. In particular, P
can cite a priority rule from precedent A, and combine it with the priority scheme
r of Section 6.3., which results in a priority argument that makes PI strictly defeat
Oi·
P~: [P2:::::} rlO/12 -< r7,
r: rlO/12 -< r7 ::::} rIO -< r7]
We now have r S = r 4, and an actual dialogue tree with still two branches: T4a =
T3a , and T4b = PI - 0i - P~.
o cannot attack this priority conclusion, and 0 cannot extend the tree without
introducing a new rule, so at this stage P is winning, since all branches of the tree
end with moves by P: so the tree is a proof that on the basis of r S the argument PI
for...., change is justified.
However, 0 can reverse the outcome by introducing a distinction: 0 can dis-
tinguish PI with the following argument.
0;': [rd14: r-.., ...., domestic-citizenship::::} ...., appl(r3/s)]
Note that rd14 has been introduced by applying Definition 5.4. At this point we have
r6 = rs U {rd 14}, and the actual dialogue tree now has three branches: TSa = T4a ,
T5b = T4b , and T5c = PI - 0;'. One of these branches ends with a move by 0 and
P cannot extend the branch without introducing a new rule, so at this stage 0 is
winning: the tree is a proof that on the basis of r 6 the argument PI for"'" change
is not justified. Moreover, since P cannot even reply with introducing new rules,
P cannot reverse this result. At first sight, this would seem to be too strong, but
here it should be recalled that our model is not intended to exhaust the heuristics
for introducing premises.
7. Summary of the Results Obtained

Let us now go back to the case-based systems discussed in Section 3, compare
these systems with our approach and discuss how the features of these systems
that we wanted to model are captured in our framework. We must state beforehand
a caveat. We have presented an abstract, logical framework and we have tried to
relate it to some features of existing systems (HYPO, CABARET, Branting's work,
and CATO): thus we have on the one hand provided an analysis of these features in
a more general and abstract theoretical framework than these systems, but on the
[ 173 ]
other hand we have ignored many of their interesting knowledge-representation

and implementation aspects, which would have to be addressed again in an imple-
mentation of our framework. Nevertheless, our abstract theory may be useful not
only to gain more insight into case-based reasoning, but also to compare different
systems and to suggest ways of integrating and developing them.
HYPO
To start with, as mentioned in the introduction, we differ from HYPO in adopting a
'normative' dialectical approach, where the aim is to test the tenability of a claim,
while HYPO has a more cognitive approach, focused more on the generation and
less on the evaluation of debates. This is reflected in two more specific differences.
Firstly, while our dialogue game has dialectical asymmetry between the parties, in
HYPO the rules are on this point the same for the proponent and the opponent: in
particular, while in our dialogue game the proponent's moves must have stronger
force than the opponent's previous move, in HYPO a plaintiff's move is also al-
lowed if it is not weaker than the defendant's previous move. Secondly, while in our
system the rule priorities are used for adjudicating between conflicting arguments,
in HYPO the more-on-point ordering is used for selecting the best move for a
player and not for determining the outcome of a dispute (with one exception).
Another general difference is that while we have a formal theory, defining the
space of possible dialogues, HYPO is an implemented program that is meant to
actually traverse such a space, i.e., its task is to generate disputes. However, it is
not difficult to imagine how our protocol, or a modified version, could also be used
to generate dialogues, by adding the appropriate control rules.
With this in mind. we can observe the following differences and similarites
between our approach and HYPO.
HYPO's case representation scheme can be captured in our framework as a
limit case, by assuming that each case contains, besides the facts, just two rules
and one priority statement. For example, we can represent a case won by plaintiff,
who argued for d, and with pro-plaintiff factors J;" 1\ ... 1\ InJT and pro-defendant
factors It 1\ ... 1\ I~, as follows.
rJT: Ir 1\ ... 1\ I:: : : } d

r8: It 1\ ... 1\ f~ ::::} -.d
p: ::::} r8 -< rJT
HYPO's dialectical moves are then captured by our notions of following, analo-
gizing and distinguishing precedents. In particular, if a case rule is analogized by
broadening a rule, then our distinction as defined in Definition 5.4 points to the
case factors that are missing in the CFS, exactly as HYPO's distinguishing does.
Note also that in our system, as in HYPO, a party can only distinguish a precedent
if the missing factor favours the other party. Note finally, that HYPO's notion of a
counterexample, i.e., attacking an analogy by citing a case with the missing factor
[174 ]
and the opposite outcome, is captured in our model by the possibility of simply
using such a case in a counterargument. In Section 6.4. this was illustrated with
the opponent's move 0 1 • There is one difference, however: unlike in our system,
in HYPO a counterexample must be more on point than the attacked case. This is
a point illustrating the difference between 'normative' and 'cognitive' dialectics.
There are more differences. Firstly, HYPO models reasoning with hypothetical
precedents, and it allows for non-boolean factors. It seems to us, however, that there
are no theoretical objections to extending our analysis with these features. We have
also extended the HYPO approach in a number of respects. As illustrated above,
our system allows for the representation of mUlti-steps precedents and correspond-
ingly for the citation of portions of precedents. Furthermore, our system allows,
unlike HYPO's 3-ply disputes, for dialogues of arbitrary depth, and our system
allows substantive debates on the criteria for comparing precedents. To expand
on the latter, although above we indicated how HYPO's more-on-point ordering
on precedents can be incorporated in our system, we have also argued that this
ordering is only one possible view on the relation between two precedents. More
precisely, if for some other reason a party regards a less-on-point case as superior
to a more-on-point case, then in our system that party can express this view by
stating a conflicting priority argument. For instance, if the proponent says that its
precedent is more on point, the opponent could counter by saying that its precedent
is more recent. Then the debate could continue on whether on-pointness or recency
is the more important standard. This is why in Section 6.4. we allowed 0 1 to state a
counterargument, although it was less on point than the analogy it attacked. Finally,
the expressiveness of our rule language, which allows for rules about rules, opens
prospects for representing teleological arguments, by which we mean arguments
referring to the purposes of rules (see also Hage, 1996, 1997). However, we leave
applications of this possibility to future research. The downside of the increase in
expressivity is, of course, a decrease in computational efficiency.
BRANTING
Let us now move to Branting's work. As we intended, we model Branting's step-

wise reduction-graph model of ratio decidendi simply by representing a case as a
multi-step argument. Correspondingly, we have preserved Banting's idea of citing
portions or precedents. A point that is not addressed by Branting is the possibility of
multiple rationes decidendi in a case (recognized by e.g. Cross and Harris (1991,
p. 81)). In our approach this can be captured by including in the case multiple
arguments for its decision.
Branting (1994) lists five evaluation criteria for models of ratio decidendi. In
Section 2 we have provided a (partial) model of the content of precedents, which
we have formalized in the rest of this article. We have also observed that the notion
of a ratio decidendi may the linked to that of (the rules in) a justified argument.
We do not want to go into the question whether our and Branting's model precisely
[175 ]
capture the notion of ratio decidendi as discussed in legal theory. Nevertheless, it

is interesting to evaluate our model in the light of Branting's criteria. We claim that
our model of Section 2 directly satisfies some of these criteria and can be easily
made to satisfy the other criteria.
Firstly, the criterion that the ratio decidendi captures the propositions necessary
to the decision can be satisfied by making sure that arguments do not contain irrel-
evant rules. Note, however, that we would say that a precedent containing multiple
justified arguments has multiple rationes decidendi, so that in our framework the
necessity criterion only applies within each ratio, not to the overall decision. Sec-
ondly, our model obviously shares Branting's observation that a precedent often
contains several abstraction steps from the facts to the decision, since it allows
for multistep arguments. Thirdly, in our model the rationale of a decision can be
grounded in the specific facts of the case by including the appropriate facts in the
justified arguments. Fourth, our account allows a precedent to contain the theory
underlying the decision, in the form of an argument with the appropriate rules.
And, finally, we share Branting's observation that the rationale of a decision can be
limited, extended or overturned by subsequent decisions, since in our framework
a justified argument of a case can be attacked, analogized and even defeated with
arguments based on subsequent precedents.
CABARET
As far as CABARET's notions of broadening and discrediting a rule are concerned,

we think that our notions of analogizing and distinguishing come close to capturing
those notions. Firstly, we allow the broadening of not just precedent rules but also
rules in CSRules, which might contain statutory rules. And CABARET's way of
broadening a rule by citing a case where a rule was applied although not all of
its preconditions were satisfied, can be approximated in our model by citing a
case that contains the broadened rule. Furthermore, our system naturally allows
the use of precedents to develop arguments concerning the antecedents of rules.
And CABARET's heuristic for discrediting a rule has its exact counterpart in our
model: an argument using the rule can be attacked by an argument citing a case
with the opposite outcome.
However, we have not addressed many other interesting features of CABARET.
As mentioned above, it would be interesting to investigate how the argument strate-
gies, moves and primitives of Skalak and Rissland (1992) could be integrated with
our model of actual dialogues.
CATO
With respect to CATO, as far as it incorporates elements of HYPO, our above

remarks on HYPO also apply to CATO. As for CATO's new features that we
wanted to include, our first aim was to be able to express a factor hierarchy. This
[ 176]
is indeed possible, by expressing it as a set of CSRuies and by expressing priorities

between these rules. We also wanted to be able to express multiple views on the
relation between factors. This is possible since our system allows CSRuies to be
inconsistent and does not assume any other property of this set.
We now come to two ofthe new argument moves that in CATO are made possi-
ble by the factor hierarchy, emphasizing and downplaying a distinction. We believe
that some aspects underlying these moves are present in our system or can be added
to it. Let us go back to our example in Section 3.4., and represent it as follows.
D: {rl/5: kept-house /\ short-duration =} --, change,
!I: kept-house, f5: short-duration}
As we remarked above, downplaying a distinction comes down to saying that while
there is a distinction at the factual level, at a more abstract level a parallel can be
drawn. Consider the following current fact situation.
CFS = U5: short-duration, !Ig: kept-old-car-number-plate}
And assume that a proponent of --, change starts a dialogue with analogizing prece-
dent D, by broadening rl/5.
PI: [r5: short-duration =} --, change,
f5: short-duration]
The opponent replies by distinguishing r5.
0 1: [rd5: rv kept-house =} --, appl(r5)]
We now make some assumptions about the content of CSRules. Suppose it contains
the following rules, formalizing a part of a factor hierarchy displayed in Figure 4.
fh l : kept-house =} showed-intention-to-return
f h2: kept-old-car-numberplate =} showed-intention-to-return
We could now define a new premise introduction heuristic for P: we could allow
him to use instead of the rule rl/5 a new, more abstract rule
r5/19: showed-intention-to-return /\ short-duration /\ =} --, change
and use this rule in a 'backtracking' response to 0 1 , with
P{: [r5/19: showed-intention-to-return /\ short-duration =} --, change,
f h 2 : kept-old-car-numberplate =} showed-intention-to-return,
f5: short-duration, f18: kept-old-car-number-plate]
Thus O's distinction is downplayed by P by 'backtracking' in the dispute, starting
a new actual dialogue tree, with an improved argument for --, change.11 This ac-
II Recall that we abstract from specific control structures for constructing actual dialogue trees
during a dispute.
[ 177]
change
I /19: showed-intention-to-retum
/1:
/~
kept-house I /IS: kept-old-car-number-plate
Figure 4. A partial factor hierarchy.
count of abstraction type hierarchy follows the infonnal analysis that was given in
(Prakken, 1993), pp. 21-2 and (Prakken, 1997), pp. 27-8.
Let us now tum to emphasizing a distinction. Consider the following current
fact situation.
CFS' = {/s: short-duration}
As explained in Section 3.4., CATO not only allows (as HYPO) to distinguish D by
saying that the CFS lacks D's factor kept-house, but it also allows to emphasize this
distinction by saying that thus in the CFS it does not hold that showed-intention-
to-return, unlike in D. Thus CATO allows explaining the difference in tenns of the
factor hierarchy. How could this be modelled in an extension of our protocol? This
could be done by allowing 0 1 to distinguish PI with the following argument.
O~: [r~s:"" kept-house::::} -, showed-intention-to-retum,
rd19: -, showed-intention-to-retum ::::} -, appl (rs)]
We leave the precise definition of this move, and of abstraction type analogy, to
future research. Nevertheless, we hope to have shown that our system provides a
suitable basis for defining case-based reasoning moves as heuristics for introduc-
ing premises into a dispute. In the present paper we have focused on two such
heuristics, viz. HYPO-style analogizing and distinguishing, but as systems like
CABARET and CATO show, many more remain to be studied.
[ 178 ]
8. Related Research on Formal Models of Argumentation
In the course of the paper we have frequently cited previous work on case-based
reasoning. However our model also builds upon previous logic-based accounts of
legal case-based reasoning. To our knowledge the first analysis of this kind was
(Loui et aI., 1993), further developed in (Loui and Norman, 1995). In the latter
paper the use of rationales in legal argument is studied in a dialectical setting. A
protocol for dispute is formally defined, and various uses of rationales within the
protocol are analysed. As we understand Loui and Norman, these uses are modelled
as ways to modify the representation of a case. Our idea to represent cases as a
collection of possibly conflicting arguments was inspired by Loui and Norman,
who use the method in formalizing a certain type of rationale of precedents, a so-
called 'disputation rationale'. Their idea is that a party who wants to attack the
use of a certain precedent, can do so by first arguing that the ratio decidendi of
the precedent was in fact the result of a choice between conflicting arguments, by
then replacing the rule of the case by these conflicting arguments, and by finally
showing that in the new fact situation the outcome of the dispute would have been
different.
Further exploiting their 1995 case representation method, Loui and Norman
(1997) analyse various ways of making and attacking analogical uses of cases.
One of them is abstraction type analogy, discussed in this paper in Sections 2.2.1
and 7. They do not embed their analysis in a formal protocol for dispute. It would
be interesting to see how their analysis can be embedded in our system as premise
introduction heuristics.
Hage (1996) gives, in the context of his 'reason-based logic' a representation
method for cases that is similar to ours in that it separates the reasons pro and con
and expresses the resolution of their conflict as a priority rule. However, his method
is not embedded in a dialectical context.
Vreeswijk (1996) has earlier embedded HYPO-style reasoning in a formally
defined dialectical protocol, but since this is not the main theme of his paper (which
is reasoning about protocol), his analysis is simpler and less detailed than ours: in
particular, he does not generalize HYPO's case representation method, he does not
formalize reasoning about priorities and he does not discuss ways of distinguishing
precedents.
Finally, we briefly compare our model with Gordon's (1995) Pleadings Game,
which formalizes and implements civil pleading as a dialogue game, thus ad-
dressing the third, procedural layer of our four-layered (and his three-layered)
model of legal argument. Both systems assume a system for defeasible argumen-
tation at the second, dialectical level (Gordon uses the proof theory of Geffner and
Pearl's (1992) system for conditional entailment). Furthermore, both systems allow
the introduction of premises during a dialogue. However, while Gordon's model
allows the introduction of any piece of information as long as it is relevant to the
discussion, in our model the only way to introduce new premises is by analogical
[ 179]
reasoning. On the other hand, the Pleadings Game does not account for analogical
reasoning with legal precedents. These differences result from the fact that while
Gordon focuses on the speech act aspects of argumentation, i.e., on the procedural
layer, we focus on its heuristic aspects, i.e., on the heuristic, or strategic layer. As
we argued several times above, we believe that a full account of legal argument
should combine these two layers.
9. Conclusion
Summarizing, we can ask what has been gained by our logical analysis of case-
based legal reasoning. We think we have gained a number of things. Firstly, we
have shown that several features of some systems developed for case-based rea-
soning in the legal domain may also be viewed as special cases of a more general
theory of defeasible argumentation, which (we hope) has illuminated and clarified
these features. This in tum has made it possible to suggest some possible exten-
sions and refinements of these systems. We also hope that our work may facilitate
the application of HYPO-style reasoning in domains similar to the law; without
embedding the HYPO approach in an abstract theory similarities between different
domains might remain hidden. On the other hand, our logic-based approach has
abstracted from many implementation issues, for which reason it is not readily
implementable.
It might be asked whether we have shown that our model is a correct model of
legal reasoning. We have not undertaken an empirical comparison of our system
with actual reasoning oflawyers (as done in Aleven and Ashley, 1997). The reason
is that our theory is, as explained in the introduction, a normative theory: it does
not aim to capture how lawyers actually argue, but to express how they should
argue. Comparing our theory with actual arguments produced by 'good' lawyers
would of course be useful, since it is reasonable to assume that many recognized
legal experts reason rationally. However, such a comparison would not provide
an adequate validation test for our theory, since then the question would shift to
whether this assumption is indeed warranted. In fact, the problem is the same as
with validating logical systems. In the philosophy of logic this is a difficult issue,
to which no clear-cut answer exists. In the present paper we have addressed it by
basing our analysis on legal-philosophical reflections, presented in Section 2.
Finally, it should be noted that our system is just beginning to address the
heuristic aspects of dialectical legal argument. At various places we have suggested
additional heuristics for premise introduction, and we have remarked that other
aspects of HYPO, CATO, CABARET and Branting's work might also be included
in our model, but we have left this for future research. We hope that the formal
theory of this paper, and its embedding in a four-layered picture of legal argument,
provides a good basis for carrying out this research. More generally, we hope that
this paper will promote future co-operation between two different sections of the
AI & law community, both interested in dialectical legal argument.
[ 180]
Acknowledgements
The authors are grateful to Vincent Aleven, Kevin Ashley, Karl Branting and
Stephanie Briininghaus for their useful and stimulating comments on earlier ver-
sions of this paper.
References
Alexy, R (1989) A Theory of Legal Argumentation. The Theory of Rational Discourse as Theory of
Legal Justification. Oxford: Clarendon.
Aleven, V. and Ashley, KD. (1996) 'How different is different? Arguing about the significance
of similarities and differences', In Smith, 1. and Paltings, B. (eds.), Advances in Case-Based
Reasoning: Proceedings of the Third European Workshop on Case-Based Reasoning, Springer
Lecture Notes in AI 1168, pp. 1-15. Berlin: Springer Verlag.
Aleven, V. and Ashley, KD. (1997) 'Evaluating a learning environment for case-based argumentation
skills', Proceedings of the Sixth International Conference on Artificial Intelligence and Law, pp.
170-179. New York: ACM Press.
Ashley, KD. (1990) Modeling Legal Argument: Reasoning with Cases and Hypotheticals. Cam-
bridge, MA: MIT Press.
Berman, D.H. and Hafner, C.D. (1991) 'Incorporating procedural context into a model of case-based
legal reasoning' , Proceedings of the Third International Conference on Artificial Intelligence and
Law, pp. 12-20. New York: ACM Press.
Berman, D.H. and Hafner, C.D. (1995) 'Understanding precedents in a temporal contex of evolving
legal doctrine', Proceedings of the Fifth International Conference on Artificial Intelligence and
Law, pp. 42-51. New York: ACM Press.
Bondarenko, A., Dung, P.M., Kowalski, RA., and Toni, P. (1997) 'An abstract, argumentation-
theoretic approach to default reasoning', Artificial Intelligence 93, 63-101.
Branting, L.K (1991) 'Reasoning with portions of precedents', Proceedings of the Third In-
ternational Conference on Artificial Intelligence and Law, pp. 145-154. New York: ACM
Press.
Branting, L.K. (1994) 'A computational model of ratio decidendi', Artificial Intelligence and Law 2,
1-31.
Branting, L.K and Porter, B.W. (1991) 'Rules and precedents as complementary warrants',
Proceedings of the Ninth National Conference on Artificial Intelligence (AAAI-1991).
Brewka, G. and Gordon, T.P. (1994) 'How to buy a Porsche, an approach to defeasible decision
making', In Working Notes of the AAAI-94 Workshop on Computational Dialectics, pp. 28-38.
Seattle, Washington.
Cross, R. and Harris, J.W. (1991) Precedent in English Law, 4th edition. Oxford: Clarendon Press.
Dung, P.M. (1995) 'On the acceptability of arguments and its fundamental role in nonmonotonic
reasoning, logic programming, and n-person games', Artificial Intelligence 77, 321-357.
Dworkin, RM. (1977) Taking Rights Seriously. London: Duckworth.
Dworkin, RM. (1985) A Matter of Principle. London: Harvard University Press.
Geffner, H. and Pearl, J. (1992) 'Conditional entailment: bridging two approaches to default
reasoning', Artificial Intelligence 53, 209-244.
Goldstein, L. (ed.) (1987) Precedent in Law. Oxford: Oxford University Press.
Gordon, T.P. (1995) The Pleadings Game. An Artificial Intelligence Model of Procedural Justice.
Kluwer.
Gordon, T.P. and Karacapilidis, N. (1997) The Zeno argumentation framework. Proceedings of the
Sixth International Conference on Artificial Intelligence and Law, pp. 10-18. New York: ACM
Press.
[ 181 ]
Hage, J.e. (1996) 'A theory oflegal reasoning and a logic to match', In (Prakken and Sartor, 1996c),
43-117.
Hage, J.e. (1997) Reasoning With Rules. An Essay on Legal Reasoning and Its Underlying Logic.
Dordrecht etc.: KIuwer Law and Philosophy Library.
Hage, J.C., Leenes, R, and Lodder, A.R (1994) 'Hard cases: a procedural approach', Artificial
Intelligence and Law 2, 113-166.
KIug, U. (1966) Juristische Logik. Berlin: Springer.
Loui, RP. (1987) 'Defeat among arguments: a system of defeasible inference', Computational
Intelligence 2, 100--106.
Loui, RP. (1998) 'Process and policy: resource-bounded non-demonstrative reasoning', To appear
in Computational Intelligence 14, 1.
Loui, R.P., Norman, J., Olson, J., and Merrill, A. (1993) 'A design for reasoning with policies,
precedents, and rationales', Proceedings of the Fourth International Conference on Artificial
Intelligence and Law, pp. 202-211. New York: ACM Press.
Loui, R.P. and Norman, J. (1995) 'Rationales and argument moves', Artificial Intelligence and Law
3, 159-189.
Loui, RP. and Norman, J. (1997) 'Eliding the arguments of cases', Submitted.
MacCormick, N. (1978) Legal Reasoning and Legal Theory. Oxford: Clarendon.
MacCormick, N. (1987) 'Why cases have rationes and what these are', In (Goldstein, 1987), pp.
155-182.
McCarty, L.T. (1995) An implementation of Eisner v. Macomber. Proceedings of the Fifth In-
ternational Conference on Artificial Intelligence and Law, pp. 276--286. New York: ACM
Press.
McCarty, L.T. (1997) 'Some arguments about legal arguments', Proceedings of the Sixth In-
ternational Conference on Artificial Intelligence and Law, pp. 215-224. New York: ACM
Press.
McCarty, L.T. and Sridharan, N.S. (1981) The representation of an evolving system of legal
concepts: II. Prototypes and deformations', Proceedings of the Seventh International Joint
Conference on Artificial Intelligence, pp. 246-253.
Mill, J.S. [1859] (1974) On Liberty. Harmondsworth: Penguin.
Pollock, J.L. (1987) 'Defeasible reasoning', Cognitive Science 11,481-518.
Prakken, H. (1993) Logical Tools for Modelling Legal Argument. Doctoral dissertation Free
University Amsterdam.
Prakken, H. (1995) 'From logic to dialectics in legal argument', Proceedings of the Fifth In-
ternational Conference on Art!ficial Intelligence and Law, pp. 165-174. New York: ACM
Press.
Prakken, H. (1997) Logical Tools for Modelling Legal Argument. A Study of Defeasible Reasoning
in Law. Dordrecht etc.: KIuwer Law and Philosophy Library.
Prakken, H. (1998) 'Dialectical proof theory for defeasible argumentation with defeasible priorities
(preliminary report)', To appear in Proceedings of the Fourth ModelAge Workshop on Formal
Models of Agents, Springer Lecture Notes in AI. Berlin: Springer Verlag.
Prakken, H. and Sartor, G. (1996a) 'A system for defeasible argumentation, with defeasible priori-
ties' , Proceedings of the International Conference on Formal and Applied Practical Reasoning,
Springer Lecture Notes in AI 1085, pp. 510--524. Berlin: Springer Verlag.
Prakken, H. and Sartor, G. (l996b) 'A dialectical model of assessing conflicting arguments in legal
reasoning', In (Prakken and Sartor, 1996c), pp. 175-212.
Prakken, H. and Sartor, G. (eds.) (1996c) Logical Models of Legal Argument. Dordrecht: KIuwer
Academic Publishers (reprint of Artificial Intelligence and Law 4 (1996».
Prakken, H. and Sartor, G. (1997a) 'Argument-based extended logic programming with defeasible
priorities', Journal of Applied Non-classical Logics 7: 25-75.
[ 182]
Prakken, H. and Sartor, G. (1997b) 'Reasoning with precedents in a dialogue game', Proceedings of
the Sixth International Conference on Artificial Intelligence and Law, pp. 1-9. New York: ACM
Press.
Raz, J. (1975) Practical Reason and Nonns. London: Hutchinson.
Raz, J. (1989) The Authority of the Law. Oxford: Oxford University Press.
Rescher, N. (1977) Dialectics: a Controversy-oriented Approach to the Theory of Knowledge. State
University of New York Press, Albany, NY.
Rissland, E.L. and Ashley, K.D. (1987) 'A case-based system for trade secrets law', Proceedings
of the First International Conference on Artificial Intelligence and Law, pp. 60-66. New York:
ACM Press.
Sartor, G. (1992) 'Normative conflicts in legal reasoning', Artificial Intelligence and Law 1,209-235.
Sartor, G. (1997) 'Logic and argumentation in legal reasoning'. To appear in Current Legal Theory.
Skalak, D.B. and Rissland, E.L. (1991) 'CABARET: rule interpretation in a hybrid architecture',
International Journal of Man-Machine Studies 34, 839-887.
Skalak, D.B. and Rissland, E.L. (1992) 'Arguments and cases. An inevitable intertwining', Artificial
Intelligence and Law 1,3-44.
Summers, R. (1997) Precedent in the United States. In press.
Taruffo, M. (1975) La Motivazione della Sentenza Civile (Justification in civil judgement). Padova:
Cedam.
Twining, W., and Miers, D. (1991) How to do Things with Rules. London: Butterworth.
Verheij, B. (1996) Rules, Reasons, Arguments. Formal Studies of Argumentation and Defeat.
Doctoral dissertation University of Maastricht.
Vreeswijk, G. (1996) 'Representation of formal dispute with a standing order', Research Report
MATRIX, University of Limburg.
Vreeswijk, G. (1997) 'Abstract argumentation systems', Artificial Intelligence 90,225-279.
Williams, G. (1982) Learning the Law. London: Stevens.
Wroblewski, J. [1969] (1983) 'Justification of Legal Decisions', In Wroblewski, J. (ed.), Meaning
and Truth in Judicial Decision, pp. 49-70. Helsinski: A-Thieto Oy. (First published in Logique
it Analyse 3 (1969)).
[ 183 ]
289
The Judge and the Computer: How Best 'Decision

Support'?
PHILIP LEITH
School of Law, Queen's University of Belfast, Belfast BT7 1NN, Northern Ireland
Abstract. One of the important lessons learned from the early expert consultants is that excellent
decision-making performance does not guarantee user acceptance. (1981)1 Version 2.1 is the final
release of ASSYST. During the past several years, the Sentencing Commission has informally sur-
veyed probation officers at training sessions to determine the usefulness of the ASSYST guideline
application software. On the whole we have found that the ASSYST program is not frequently used.
After balancing the program's usefulness with the resources required to keep it updated, we have
decided to discontinue maintaining ASSYST. (1996)2
1. Introduction
From a period in the early 1980's much of my research effort was spent in arguing
against the utility of legal expert systems and legal logic. 3 The effort ceased with
the publication of my Formalism in AI and Computer Science 4 which argued that
the representational formalisms being used by those in AI and computer science
were flawed - that they could not, for example, represent the richness of legal
knowledge in any useful way. For most of the 1980's my perspective was obviously
flowing against a strong tide: logic programming was in the ascendancy and AI
still seemed to offer much in terms of practical legal expert systems. However, as
I shall make clear in this paper and the quotations above suggest, there has always
evidence in the background that there were and are substantial problems with the
attempts of AI researchers to persuade users to make use of their systems, even
from before the explosion of interest in the application of expert systems. The
1 E.H. Shortliffe, 1981. Shortliffe was the primary designer in the MYCIN medical expert system.
See also Shortliffe, 1995; Shortliffe, 1994; Shortliffe, 1992. Shortliffe has been one of those who has
been most considerate of user problems in AI. His field has been medicine, but there is little doubt
that the discretion given to doctors is relatively similar to that given to judges.
2 ASSYST web page at www.uSsc.gov/assyst.htm.This is a rule-based advisory program related
to the US Sentencing Commission's task. For a description of this in AI terms, see Simon & Gaes,
1989.
3 See for example, Leith, 1986a; Leith, 1986b.
4 Leith, 1990.
[ 185 ]
290 DR. PHILIP LEITH
thesis of this paper is a restatement of this argument: those who wish to provide
judicial decision support systems would be well advised to heed any lessons which
might be learned from the problems of user acceptance.
It may be that there are many 'expert systems' which are presently in day-to-
day usage. However, given the fact that there is a distinct lack of confidence in any
definition of what an 'expert system' is and indeed which programs demonstrate
'artificial intelligence', any discussion of the success of AI and expert systems in
general is difficult to conclude. In law though, the situation is clearer: all those
'expert systems' which have been built in the academic environment have not
moved successfully over to the workplace. s Many in AI will not agree with this, but
it is a firm position I take which has not yet been proven incorrect by the evidence
available - see the Appendix to this article.
In this paper, I want to revisit the arguments against what I described as 'for-
malist' approaches and suggest that the current debate about 'decision support' is
simply the rerunning of old approaches which were found in the expert systems
movement of the 1980s. I shall discuss two approaches, the cognitive and the
sociological which can be used in analysis of the judicial process. These are dia-
metrically opposed. One emphasises the techniques which model human processes
(i.e. the 'artificial intelligence') whereas the other emphasises trying to understand
the social context of the problem. The context of the problem, I suggest, is much
more important than modelling human processes. My general argument is that in
order to do computing and law properly - whether it be inspired by AI or by tra-
ditional data-processing techniques - the most important first step is to understand
the problem. If one understands the problem and the problem is solvable, then
deciding which technique to use becomes relatively trivial.
2. What is 'Decision Support' - Just a Failure of the Expert Systems

Movement?
Informally, it is easy to see what the concept of 'decision support' means. It is an
attempt to support decision makers in their day-to-day tasks. However, we are all
essentially decision-makers (e.g., deciding whether to take an umbrella to work
or not) and there are many such daily decisions which must be made. There is,
therefore, something apart from this general support for decision making which is
implied. It is:
1. Decision making in a complex environment, and
S A consistent position taken by AI researchers is to suggest that there are actually working legal
expert systems, and therefore that my assertion is incorrect. I deal with this below in the paper's
appendix, suggesting that such claims to success should be viewed with more scepticism, and that
for those programs which seem to have moved over are usually built with pre-AI techniques and aims
(e.g., Juricas from Rotterdam). Typically one hears anecdotal evidence which is not too far from, "I
was at a conference and spoke to a friend who had a cousin who once saw a program which was an
expert system, and my friend's cousin said ... ". I exaggerate, but only slightly.
[186 ]
THE JUDGE AND THE COMPUTER 291
2. Decision making which is based in scarce 'expertise'.

These elements are, of course, basic to the idea of the 'expert system' (but do not
fully define it) as it was developed in the 1980s and, it can be argued, the concept
of decision support system owes its existence to that of 'expert system'. It is not
clear from the literature, though, just how or why the idea of expert system was
transformed into the idea of 'decision support system'. This transformation is a
limiting of the idea of expert system. But what aspects of an 'expert system' are
not to be found in the idea of 'decision support system'?
If we look back to the history of AI we can, perhaps, try to find out how
the transformation occurred. The lack of success of the General Problem Solver6
model of AI where general solutions to learning and problem solving were to be
found gave rise to the interest in complex environments, since these environments
appeared to be relatively closed off from the 'noise' of general cognition. The
laboratory success of programs which were produced in the late 1970s such as
DENDRAL, MYCIN and PROSPECTOR appeared to demonstrate that complex
environments might actually be easier to handle than more diffuse domains. That
these programs offered expertise which was rare and expensive was seen to be a
distinct advantage since it promised that AI would be a useful discipline. The 1980s
then became, as we all know, the decade of the expert system.
In the laboratory it is clear that some programs which were being designed
could be used in a predictive manner. They did, as the terminology of the decade
suggested, 'reason like experts'. Certainly MYCIN was the most heavily discussed
of the early programs and there is no reason to suggest that - in the very small area
in which it worked - it could not help decide which antimicrobial drug should be
given. Many other programs were less well analysed. Indeed the zeitgeist of the
time seemed to be (as can be seen from the papers at AI conferences) that the re-
searcher had to describe what his program was going to do and how intelligent and
useful it would be even before it was programmed. Programs were also frequently
heavily discussed without, it appears, any link to reality. For example, Duda, a
member of the PROSPECTOR team (which was never actually used in practice)
wrote:
The widespread interest in expert systems has led to numerous published
articles that cite impressive accomplishments - practical results that justify
much of this interest.
Having been primary contributors to the PROSPECTOR project, we are
particularly sensitive to comments about that program, such as the one that ap-
peared in a recent book review in this journal that referred to " ... PROSPEC-
TOR's hundred million dollar success story". Unfortunately, this particular
statement, which is similar to others we have encountered elsewhere, has no
factual basis.7
6 Newell & Simon, 1972.
7 Duda, 1985.
[ 187]
DENDRAL's team even suggested that the lack of feedback they got from users
was an indication of successful use (rather than - as most producers of programs
would realise - a lack of use of the program) by writing:
Many persons have used DENDRAL programs ... in an experimental mode.
Some chemists have used programs on the SUMEX machine, others have
requested help by mail, and a few have imported programs to their own com-
puters .... Users do not always tell us about the problems they solve us-
ing the DENRAL programs. To some extent this is one sign of a successful
application. 8
And MYCIN, too, when it was moved over onto the hospital wards went unused.
It is important to concentrate clearly upon this history. There have been many
hundreds of programs which have been produced as 'expert systems' and which
grew from the optimism created by these three programs, MYCIN, DENDRAL
and PROSPECTOR. Though early, they were important in defining the mood and
direction of much work in the 1980s. But, did these programs make the move from
the laboratory to the world of user? My own belief is that these programs were not
successful in their move from the labs: and, others, following through the lists of
working expert systems, have found little truth in the claims9 of 'real-life' use.
Though some in AI and law continue to argue for the success of the expert
systems concept, it is clear that many researchers have moved on and accept that
there has been a failure. Case-based reasoning is an example of the move away from
expert systems in the 1980s mode (Le., heavily rule-based) towards one which is
much, much closer to the Jurimetrics lO model of analysis of cases and prediction
according to factors in the case. With case based reasoning, exemplars are analysed
by hand and coded according to the existence of factors deemed as potentially im-
portant - these are then used to compare the case in hand. Just why such techniques
are described as being based in AI is difficult to ascertain, unless one sees them as
arriving through the history of expert systems.
Though the term expert system has been abandoned by many, the cognitive
flavour of that type of research remains: the aim to replicate the reasoning processes
and thinking of the expert. Though this cognitive flavour remains, many researchers
now prefer to talk about 'decision support' being their aim, rather than the creation
of an artificial expert.
What has actually happened here? Is there a conceptual move? It is certainly
not clear from the literature, but I suspect that what has happened is that the early
realisation by, for example, Shortliffe that the difficulty in the research work lies
in getting users to want to use systems has been slowly gaining ground. The expert
system, providing its 'immortal' knowledge to non-expert users (as was claimed
in the first batch of 'popular' literature) has never been a go-er (particularly in the
8 Buchanan, 316, 1981.

9 See for example, Ostberg, 1988.
10 See, Baade, 1963.
[ 188]
legal area). The second best goal, then, has been to provide neo-expert systems
which are directed at experts. And, in order to put distance between the earlier
failure and the future, hoped for success, the terminology has been changed to
'Decision Support System'. But will change in goal and terminology be sufficient
to ensure user acceptance?
3. Problems with the Cognitive Model

The cognitive model which has driven AI has been one which is intrinsically psy-
chological and individualistic. From the earliest days of AI when it was involved in
- for example - GPS, the attempt has been to locate the area of interest in the think-
ing processes of the individual. When expert systems arrived, this cognitive model
continued: rule representation was seen to be the means of using and explicating
knowledge - for example, when an expert system was said to have an 'explanatory'
mechanism, what was really meant was that it could list the rules which had been
triggered. Explanation was seen to be wrapped up in the model of rules, which is a
psychological model.
There are obviously advantages in this psychological model. It is relatively clear
what the object being targeted is - professional problem solving, say. It is also
closely linked with a highly developed experimental discipline (psychology and
its subset, cognitive psychology) so that - if need be - there is much expertise
in testing and designing empirical research. And, given that AI has been closely
linked to the computational metaphor during its history, the benefits of having
dialogue with the wider psychological movement should help overcome a potential
blinkering effect of the computer - that is, seeing everything from the perspective
of how models can be computerised.
In the event, the advantages which might have accrued from AI's psycholog-
ical model have not been particularly evident. The targeting has been, perhaps,
too focused with an extremely unbalanced view of, as in this case, professional
knowledge. AI has also not been particularly adept at borrowing from the empir-
ical expertise of psychology: the only experiment has been to produce a program
and suck it and see; a 'sucking and seeing' which has sometimes been highly
superficial. And, the computational model has not been ameliorated in any way
by the contact with the broader field of psychology. In many ways, there has been
antipathy between mainstream psychology and those in AI.
But even if AI had paid more attention to wider psychological approaches and
techniques, would it have made much difference? I think not. It seems to me that
the problem which would not be answered by the psychological approach would
be that of defining the context of a computer problem, so that a solution can be
achieved. This is primarily one of systems analysis, rather than psychology. The
systems analyst has to look at the wider context of user need, function and role
before any proper decision can be made about how best to design a program. The
psychological approach doesn't do this. Rather it takes as a given the individual
[189]
and tries to understand/model what that individual does. When applied to AI, this
would mean that after analysis and repackaging of 'knowledge' or whatever one
wants to call it, the expectation is that individuals will want to use a representation
of the original behaviour pattern. Why should they? Has there been any stated need
for this? And, may there not be another way to better achieve the same aim without
simply remodelling one behavioural mechanism II?
We have much evidence that the professional has much control over how he
fulfils his tasks. We can see from the literature on implementing judicial or court-
based systems that there has been a particular context which has allowed this to be
supported. Most frequently, it appears that the reasoning has been that the judges
have seen political advantage in helping develop a 'decision support system', yet
when the system has been produced, have made little use of it. There is certainly
little evidence that the jUdiciary are actively looking for decision support systems
which will help them to carry out their functions, and given this lack of desire we
should be careful before we begin to assume that judges will accept these systems.
4. The Problem of Judicial Discretion

Since discontinuing writing about AI and law at the end of the 1980's, most of
my research has been into the practice oflaw (and also the use of computers in that
practice I2 ). I have looked at lawyers in the UK I3 as well as international practice of
patent law. 14 The closer I have looked at lawyers in practice, the less useful do most
of the AI programs appear. They do not seem to direct themselves to the particular
problems of being a lawyer or being a judge and, at worst, appear irrelevant to
legal practice. 15 There is, of course, now a huge literature on lawyering and judging
which should give insight to researchers into just what it is that they are trying to
achieve, but one rarely finds this literature referenced in the research into AI and
law. If, however, my assertion that the real problem in designing 'decision support
systems' is one of understanding user needs, it seems to me impossible to produce
usable programs without understanding the actual legal process as users experience
11 And there is little discussion in the AI literature of which technique is most appropriate -
traditional or 'knowledge based' - with many problems seemingly shoe-homed into an inappropriate
AI technique. This is not a problem if the aim is development of techniques; however, it may be
one if the aim is user acceptance. The conclusion from work on the rule-based Nervous Shock
Advisor (Deedman & Smith, 1991), for example, was that "While these weaknesses [which arose
from using rules] could be palliated to some extent by good rule-base design, they could not be
overcome entirely in a system that relied on rules to do everything".
12 Leith & Hoey, 1998.
13 Morison. & Leith, 1992.
14 Leith, 1998.
15 Deedman and Smiths' Nervous Shock Adviser for example, was developed by a legal prac-
titioner but seemingly not a practitioner in this area. Rather, it was chosen because "it seemed
manageable". See Deedman & Smith, 1991. The article appeared in a text called Operational Expert
Systems in Canada. Note that while, despite the title of the collection, the program was operational,
it was only so in the laboratory.
[190 ]
it. Why do AI researchers believe that success can be achieved without this research
link to practice?
In effect, I am arguing that a sociological approach should be undertaken by
those involved in AI and law. If they understand the context of the legal process,
then the techniques they advocate can be used successfully. If not, their programs
will never make the transition from the laboratory to the courtroom.
One of the major problems in designing decision support systems is that the
courtroom is a place with much discretion. As recent political debate in the UK
and US over minimum sentences etc. shows, discretion is valued highly by judges.
Any computerisation project must acknowledge that this discretion exists, and that
it is intrinsic to the task of legal adjudicating. Failure to acknowledge it will lead
to computer systems being viewed as too formalistic by users and, if labelled in
this way, they will certainly not be used. The sociological approach takes this
element of discretion as fundamental because it is also a fundamental assertion of
a sociological approach to law that the most important element in the legal process
is the actor. 16 To positivists and others, the fundamental unit is the legal rule or the
case or the case and the rule. 17 In the sociological way of doing things, then, rules
and cases are not ignored, but are simply seen as artefacts which are produced and
handled by actors. The actors are the important element in the situation, because
they are capable of interpreting rules and cases in diverse ways, according to the
context in which they are to be found. This implies that if we want to handle
law by computer, it is a mistake to concentrate upon the rule (or case) since the
fundamental source of the rule (or case) is its social use - i.e., whether or how the
actor decides to use it or not.
It also implies that since social usage is not necessarily formalistic, it is not
possible to say that rules (or cases) can always be represented in any formalism.
They may be, and then again they may not be. Some legal rules (or cases) are
applied in a highly formalistic manner. We can think of car parking offences which
- no matter how unjust the driver may feel in a given circumstance - the system is
set up simply to process and fine whoever is ticketed. Attending court and trying to
argue that the fine should not be paid will simply be seen as time-wasting. Rarely
is discretion allowed. It is almost as though the car owner is expected to take his
punishment 'like a man' and pay up -legal liability here is seen as extremely strict.
It would be easy to produce a program which modelled this non-discretionary,
strict liability system. Unfortunately, for most legal actions, the court system is
not so lacking in discretion. These are the ones which are interesting and which
should be subject to discussion. These are the ones where facts and rules and cases
are sometimes agreed and sometimes disputed, and where the basis for decision
is necessarily non-formalistic. Prediction in these situations may be statistical, but
cannot be made according to a formal knowledge representation.
16 In the meaning of 'one who performs an action'.

17 There has been a debate in AI and law on whether rule-based systems alone, or case-based
systems alone are the way forward, or whether there should be a combination of the two approaches.
[ 191 ]
How much evidence exists for suggesting that this problem of discretion is part
and parcel of the system - indeed is at the heart of the system? I argue that there
is much. In the following sections I shall look at the UK court system. I use the
UK system for convenience, but there is a strong argument that most of the aspects
of legal process in the UK can be found in other jurisdictions, too. And where
the details may match exactly between jurisdictions, the general complexity of
situation is, I have no doubt, common to all jurisdictions.
5. The Courts and the Judiciary

There is a substantial amount of investigative material on the court process. From
the early studies by ethnomethodologists where a tape recorder was placed in
the jury room, through statistical analysis of decisions, to ethnographic studies
of courts, we have a huge bulk of material which argues a general point. This
point is that, in practice, the legal system differs markedly from that proposed by
the traditional 'black letter' law view - one which suggests that legal process is
fundamentally rule-based l8 and operates according to strict precedent. There are
a number of points of divergence, which to a large degree depend on the position
in the hierarchy in the system: the lower down the court scale one goes, the more
is there a concentration upon speedy and efficient processing of matters; and the
higher up the scale, the more concentration upon esoteric questions of law. The
'black letter' view has - as many commentators have suggested - concentrated
upon these upper echelons (the Court of Appeal and the House of Lords in the
UK) and neglected what has been happening in the lower courts. Yet despite - or
because of - this lack of academic focus, it has been the lowest courts who have
managed to make use of information technology - the first courts in the UK to use
computers were Magistrates' Courts, for instance.
It is not only that the traditional view of court practice itself has been under-
mined by these studies, but also we have been provided with a way of looking at
what happens to cases before they arrive at court. Thus we have seen studies of
the various kinds of 'plea-bargaining' which occur: where defence and prosecution
agree lower charges, say, in return for a guilty plea etc. And in the UK, over 85%
of criminal cases are processed with a guilty plea. That is, only 15% of criminal
cases will involve judicial decision making which is not to do with sentencing.
The anecdotal evidence we have from civil cases indicates that of every 100 cases
initiated, 10 will go to the courtroom doors and only 1 will be fully considered by
the court. This indicates, perhaps, that while the judicial role is important (since
it overshadows all the preliminary work on cases and decisions as to whether to
proceed) in terms of the importance of judicial decision support to the legal system
as a whole, it is not over-riding.
These studies have all led to an appreciation of the amorphous nature of the
legal process - that much of the relevant decision making happens out of sight
18 And that these rules of operation are closely akin to formal rules, rather than behavioural rules.
[ 192]
of the judge. Indeed, there is much concern that this process acts against wider
concepts of justice: victims and relatives of victims lose their retributive reward
when the defendant can successfully plea-bargain (frequently outwith the view of
the judge) for a much lesser charge than victims believe to be just.
However, for consideration of judicial decision support systems, we have two
clear areas which most AI researchers believe are of interest: that of sentencing and
that of consideration of legal rules. The questions must be, just how formalisable
are these two areas in practice; and are they viewed as needing support by their
potential users?
In the next sections, I want to look at studies into the UK legal system, which
have particularly concerned themselves with judging. The arguments from each of
these suggest that the nature of judging does not accord particularly well with the
traditional view of the judge as a rule- or case-prescribed 'machine'. This will have
consequences for my argument on just how possible are decision support systems
in justice administration. This argument is that, in order to insert decision support
systems into the adjudication process, there is a requirement for two basic factors:
I. There must be a clearly felt need for decision support;

2. There must be a desire by potential users for the information provided by any
decision support system;
And further, since all computerisation involves a formalising of tasks into an or-
dered and coherent 'system', that:
3. The design of a decision support system should not detract from the discre-
tionary elements which are at the heart of notions of 'justice'.
These requirements have not always been met by many of the AI and law programs
produced. 2o In many ways, this has been because there has been an element of
pioneering or experimentalism about these projects. The aim has perhaps been
more to see how far AI techniques could be pushed in the direction of law, rather
than how successfully they could be implemented in practice. This is fine, so long
as the experimental nature is realised and clearly stated. However, if our aim is to
produce working and usable systems, such a methodology is not sufficient. Ifwork-
ing systems are desired, then - as the evidence from software failure demonstrates
- we have to think carefully about the needs of users.
My assumption is that computerisation involves formalisation. It is clear that
much of the research work in the 1980s accepted that formalisation was a requisite
for successful operation of systems: Ronald Stamper's LEGOL 21 tried to provide
a logical language for law; logic programming attempted to translate legislation
into hom clause formalism; and, frame-based formalisms were seen as essential by
20 Though some, e.g., Hutton, 1995 have begun to consider the problem in a much more coherent
manner than it has been approached in the past. Whether they see themselves as promoting AI is a
moot point. My reading is that they are closer to the sociological approach I advocate.
21 See, for example, Stamper, 1980.
[ 193 ]
others. These attempts were not particularly successful, and some have suggested
that it is possible to computerise without the degree of formalisation which was
considered essential in this previous work. For example, Prakken, 22, Sartor23 and
Gordon 24 have been proponents of a view which suggests that logic can be used
in a less positivistic manner.. Prakken has criticised the critics of earlier logical
research who suggest that it is necessarily too positivistic. His argument is that
though we can use logic to represent knowledge in a logical language, it is not
necessary to be bound totally by this logical representation. That is, that other
processing techniques (analogical, inductive or abductive reasoning) can be ap-
plied to this representation, so that features such as understanding rules and cases
within a context can be gained which are not available under the logical approaches
used in the 1980s. These techniques, for example as in the work of Prakken and
Sartor,25 can be used to provide tools to encourage reasoning about law by helping
construct arguments and also putting into doubt the arguments of an opponent. In
effect, if this approach is successful, that logic is being used in a highly social and
non-formalistic manner.
I have doubts about this approach. Logicians have, for long, tried to escape the
deductive power and/or constraints of first order logic ('fuzzy logic' was an attempt
to do this). In logical circles, these attempts have met with much criticism relating
to their technical basis (for example, attacks that the approaches are 'not really
logic at all' - see Haack for more detaiI 26 ). However, on a more intuitive level
we can see that such approaches may be useful: the simple act of thinking about
something usually helps us to clarify and resolve problems.
More importantly, though, for my argument here is the prosaic fact that we have
no real evidence that such non-formalistic approaches actually work in practice
- no system has been built operating upon these principles which is being used
by lawyers. It may be that such systems are technically possible and needed by
lawyers, but we have no evidence of this. In fact, it may be suggested that lawyers
do not really need a system to help them construct argumentation models at all:
they seem to be perfectly well able to do so without such tools. Further, I may be
wrong in this, but my hunch is that such systems would still be too formalistic for
practical use.
My essential assumption therefore remains: computerisation imposes a formal
structure upon procedures and knowledge which detracts from the social elements
(particularly discretion) which are essential requirements of certain judicial tasks.
Other tasks can be formalised without problem (court record keeping, for example)
or opposition from potential users, but these are not the judicial decision support
tasks being discussed in this paper.
22 Prakken, 1997.
23 Sartor, 1997.
24 Gordon, 1995.
25 Prakken and Sartor, 1996.
26 Haack, 1978.
[194 ]
5.1. THE MAGISTRATES' COURT

The magistrate's court in the UK is the lowest level of criminal court. It copes
with a multitude of cases which deal with motor vehicles, petty crime, drunken
behaviour, assault, etc. In an important manner, the Magistrates' Court is both a
court for dispensing justice, but also for imposing social control. Frequently the
judges in these courts are not legally qualified (but they will have a clerk who is
legally qualified27 and who will always be on hand to provide advice to the judges).
The Magistrates' Court, in terms of numbers, deals with most cases. It has been
the court which has been able to use Information Technology most successfully,
in part because there is a processing requirement much like that of the electricity
supplier. 28
In practice it has been the case that sentencing has been diverse and problem-
atical. Evidence for the differences in sentencing patterns in different courts have
caused public outcries at various times. However, there has been limited interest in
achieving a harmonised system, and there has certainly been no clear call from the
magistrates themselves asking for help in making sentencing decisions. According
to one Magistrates' Clerk:
It is difficult to see how fundamental change in an area's sentencing practice

can be affected unless a substantial number of the magistrates meet and agree
upon it. That prospect is fairly remote. ... The individualised approach to
sentencing is very firmly entrenched indeed. Even if "uniformity of approach"
were achievable in the magistrates' court, we would still need to ensure that
sentencers attached the same degree of importance to all the important vari-
able factors or sentencing disparities would continue. 29
Turner further points to a survey by the Magistrates Association which suggests

that guidelines in sentencing are used by magistrates, albeit with substantial al-
terations in their application - 129 of the 195 surveyed indicated that they made
such changes. However, the study was not particularly academic and rigorous in
design and, as Turner suggests, "it seems that guidelines are widely used, though
how they are applied in practice is another subject requiring investigation". Turner
suggests that there is actually a different form of adjudication being applied in the
courts: traffic offences are being treated strictly according to guidelines, but the
other instances which come before the magistrates are dealt with in a much more
individualised manner.
It seems clear, therefore, that though there is a concern about the evenness of
sentencing, there appears to be no desire to have a computer system which helps
magistrates to achieve sentencing consistency in all areas. Indeed, Bainbridge's
27 Though the legal qualification can be minimal- see Turner, 1992, 198.
28 See Adams, 1972 for an early view of this. There are now a variety of sources of information
on this way of viewing courts.
29 Turner, 1992: 199/200.
[195]
CASE system30 which was designed to achieve just this kind of support for mag-
istrates went unused. Magistrates are certainly prepared to be guided in a broad
manner in most cases, but in some cases they do feel obliged to follow rule-based
behaviour - that is, traffic offences. However, it is clear that there is now a move
away from using magistrates to adjudicate in many traffic cases - we have seen
in the UK the introduction of fixed penalties so that the magistrate is becoming
less involved in this kind of decision. Perhaps this raises the interesting question
of whether, when a criminal activity is the subject of highly formalised sentencing
and strict liability, it is actually necessary to use magistrates justice at all? In effect,
when the situation become ripe for computerisation, it is no longer a properly
judicial activity.
But for those who have looked at sentencing from the psychological perspective,
the complexity is apparent but there are patterns which do show that discretion is
attempted in a meaningful way. For example, Lawrence concludes:
In summary, these analyses reveal the complexity of the network of influences
on sentences for a common offence like shoplifting that magistrates mostly
see as less serious than drink driving. There is clear evidence of disparity
of treatment at all levels of the analyses of the sentencing patterns, yet the
patterns do not appear to involve capricious or arbitrary punitiveness. They
seem to be genuine attempts by magistrates to achieve sensible sentences. 3 !
For those researchers who are concerned with the area of legal rules and cases,
rather than sentencing decision support, the evidence we have of the rule-governed
behaviour of magistrates suggests that it, too, is problematical for decision support
design. In a classic early study of magistrates, Pat Carlen carried out ethnographic
study of two courts. The resultant perspective presented is one where the magis-
trates have a proclaimed rule-governed system, but in fact the rules are used in
a more problematical manner. Of the rule system which the magistrates publicly
utilise, Carlen suggests it is presented as a form of 'legal absolutism' (and note
that the legalism is applied by those who are frequently without legal training or
qualification):
The substantive dimensions of this legal absolutism portray legal processes as
being homogenous, unproblematic, external, inevitable, essential and eternal.
Strategies by defendants which threaten to reveal legal processes as being
pluralistic, problematic, intentional, contingent, phenomenal and temporal are
suppressed by remedial routines comprised of strategically interlocked so-
cial rules whose tacit components, though suppressed, are treated as being
absolutely judicious. 32
But in practice, applying legally absolute rules is not the task of the Magistrates'
Court. The court has the important element of restating society's goals and aims
30 See, for example, Bainbridge, 1991.
31 Lawrence, 1994.
32 Carlen, 1976: 128-129.
[ 1961
- that is, in enforcing social control upon those whose goals and aims are not
necessarily those of the 'larger society'. For example, drug users do not always
follow the reasoning that cannabis use will lead to heroin and the breakdown of
society; not all motoring offenders are convinced that society's needs are met by
strict liability in, for example, the matter of defective lighting; and drinkers don't
always feel gUilt about being drunk in public. Accepting the validity of these views
would undermine the very system which the magistrates are attempting to impose
upon these 'deviants'. As Carlen suggests:
In legal settings and in other settings more or less serving as agencies of social
control, the legitimated goal of theory is not to reveal an alternative, possible
society; questioning the inevitability of the social reaction to deviance threat-
ens to do just that. The only discretion to which magistrates will admit is to
that accorded them by the formal rules of law. In cases where magisterial dis-
cretion is formally prescribed and invoked, implications that such magisterial
discretion is open to situational negotiation or legitimation are indignantly
repudiated. 33
Carlen is arguing that there is a very significant area of discretion open to the
magistrates. They, however, in order to make their task of social control easier,
deny this discretion to be anywhere near its actual reality. Instead, they suggest
that the formal rules govern their behaviour entirely. In reality, the formal rules
play little part in the decision-making of the magistrate: the defendant in the court
must be objectified and squeezed into the relevant classification, in order to be
speedily processed, that the court can get on with the next business. The squeezing
is carried out by the various professionals involved in the process - court staff,
probationary staff and police prosecutors, as much as by the judge. Those who
have been onlookers at the Magistrates' Court will certainly have sympathy with
Carlen's analysis.
What should designers of decision support systems learn from these kind of
studies of magistrates? Perhaps that systems which provide advice on sentencing
and rules might be used in a more complicated environment that is first imagined.
5.2. JUDGES IN THE HIGHER COURTS

From a variety of projects we know that judging is an isolated activity - courts are
becoming larger in order to become more efficient and the nature of the task is that
the judge rarely sees other judges in operation. One study of criminal judges in
Philadelphia suggested:
Nearly all judges interviewed were surprised by the lack of collegiality among
their fellow members of the court. There was almost no socializing outside
of a few official functions. Each judge seemed to have two or three close
33 Carlen, 1976: 120.
[197]
friends among his colleagues but was rather disdainful towards most of the
other members of the court. It was surprising to have every judge interviewed
comment so negatively upon the ability of his colleagues. The most typical
descriptions involved the terms lazy, slow-witted, inadequate, and mediocre. 34
Rock, in his study of an English Crown Court, too, commented upon the difficulties
judges have in communicating their ways of handling problems. The court studied
was relatively small, and allowed communication in a way that larger courts did
not. One judge in the study suggested that this meant it was possible to:
know how other people work and how they are thinking. It helps you and
stops you doing what used to be the very bad habits in old one-judge courts of
developing idiosyncrasies that you don't have the opportunity of noticing. A
judge who comes into lunch and says, 'Guess what I did this morning?' - he's
done something rather outrageous and the others say, 'What? You can't have
done that!' It's very valuable. 35
It seems relatively unproblematic, then, that judges are keen to find out just what
their fellow judges are getting up to: whether it is with the intention of improving
themselves or substantiating their negative attitudes towards the others. Communi-
cation, it seems, is a good thing. In one way, the move towards sentencing decision
support systems appears to be such a form of communication, and might be seen to
be welcome. Unfortunately, at all levels of the courts, we do not see evidence for
this being seen in a positive light at all. All the information points to the fact that
where communication is seen to be one-way (i.e., 'top down'), then the judiciary
are opposed to it and see it as negatively impacting upon their judicial discretion.
The history of the separation of powers in government has always been prob-
lematic. The judiciary, under this philosophy, are seen as being merely technical
implementors of the desires of the legislature. This, of course, has never been the
case36 and there has always been the potential for a battle of wills between the
judiciary and the legislature and executive. Sentencing has been a recent example
here (with the Sentencing Commission in the US, for example 37 ) but there have
been other problems, too. For example, the attempt to improve court productivity
through judicial case management (this is a factor in the Woolf Report 38) has been
longer lasting in the US than in the UK and has been the focus of unhappiness
from the judiciary. Sheskin and Grau in their analysis of the attempts by the Ohio
Supreme Court to enforce court productivity through new rules of practice on the
34 Wice, 1981: 165.
35 Rock, 1993: 183.
36 Indeed, one of my arguments against logic being able to formalise the British Nationality Act
was that it included an 'ouster clause'. See Leith, 1986.
37 See Tonry, 1992.
38 See Woolf, 1996. Currently in the UK there is an attempt to move more decision making in the
Magistrates' Courts over 10 the Clerks (who welcome this) for reasons of efficiency. This is opposed
by the Magistrates who argue that it is removing judicial functions over to the administrative sphere.
See the report in The Times, 10th February 1997.
[ 198]
judiciary, suggested that the lower court judges had tactics available to undennine
these attempts:
although their professional norms precluded collective actions to oppose or
change the rules, they were not without defenses. The rules undermined their
professional status, but this very status was a weapon against those who sought
to rationalize their work. Respect for judges' professional status prevented the
Supreme Court from publicly embarrassing or punishing those who failed to
meet the established guidelines. Nor did administrative officials challenge the
judges' professional integrity by auditing their statistical reports ....
Faced with exploding demands for their services and dwindling budgets
to provide them, the courts confront a dilemma to which there is no simply
solution. No longer free to expand the quantity of services, they must develop
new, less costly resources as well as make more 'efficient' use of old struc-
ture and functions. But this transformation is not a simple structural reflex to
changed circumstances. Rather it is a process impeded by the interests and
power of those upon whose work administrative reforms impinge. 39
The general conclusions which we should draw from these kinds of researches are
that though there are problems in judging which the judges themselves are well
aware of, the solution is not simply a case of providing some technocratic solution
and expecting it to work. The judiciary have historically been highly sensitive to
attempts to undennine their discretion, and it must be assumed that any attempt
to impose computer-based decision support solutions upon them will simply not
work. Hutton et al. are clear about this, when they try to produce a system which
is not prescriptive, but descriptive (although there was also clearly a desire to
be prescriptive and 'reform' sentencing without, they believed, affecting judicial
discretion). They were aware that a threat might exist:
there was a concern that some judges might perceive the prototype Sentencing
Information System as a threat to their independence. Not only might this lead
to resistance in the designing of the prototype, but even if resistance could
be overcome, judges might be reluctant to consult a Sentencing Information
System. 40
Other researchers - particularly those from AI - have been less receptive to these
problems, which can certainly undermine the viability of a project.
If sentencing has been a particular problem, what about the use of rules? Are
these to be found to be less problematic when we move up the court hierarchy? The
evidence is that they are not less problematic. Certainly in the more esoteric areas
of legal practice where law is being fought and constructed by public agencies
and private corporations, there is much battling over legal interpretation. The High
Court, the Appeal Court and the House of Lords are surely locations where much
39 Sheskin & Grau, 1981: 248-289.
40 Hutton et aI., 1995.
[ 199]
difficult analysis of precedent is undertaken and the balance of precedents must

be considered. We see this in the Law Reports. However, these are not really the
agents to whom most AI researchers have directed their activities. Of more concern
has been the general practitioner, rather than the esoteric expert - the general prac-
titioner is clearly the potential user of case-based reasoning system or rule-based
expert systems. Do these general practitioners actually require such systems? It
appears that there is little support for the argument that they do, certainly in the
UK. We know how little use is made of legal information systems and how rarely is
'academic-style' legal research carried out by the typical practitioner. 41 In my own
research into the professional life of barristers 42 it was clear that most provincial
barristers (i.e., the vast majority of barristers) are not particularly involved in this
kind of research. One commercial barrister suggested the typical view that law was
important, but it was not necessary to carry out legal research:
lots of [other barristers] say you never have to use the law you learned at
university but thai's not my experience .... I use quite a lot of law, largely the
kind of basic law that one learned at university ... you have to make quite a
bit of reference to contract, equity and that sort of thing. (94)
And one very successful barrister suggested:
You don't necessarily have to go into much detail. The sort of opinions I'm
delivering, they're adequate; they're a fairly generalised view of matters ...
enough to allow the clients to make decisions ... Cases where you have to read
all the authorities and go into all the arguments are fairly few and far between.
(94)
And one arguing that the basic need in a barrister is to be practical, spoke about an
academically inclined colleague:
. .. there is a completely different sort of barrister to the one I hope I am ...
the one who enjoys playing with the law .... Funnily enough we've got one
in chambers ... he loves talking technical points, and all the judges know it
... he's a bright guy ... he knows ten times as much law, a hundred times
as much law, as I will ever know but I don't believe he's as practical in his
approach and that I believe causes problems .... Who are you serving? . .. At
the end of the day, technical problems can usually be overcome, You're only
delaying the evil day in most cases. (95)
Of course, this is the barristerial view, but barristers are highly attuned to the needs
of judges:
Well, county court judges, if they see someone with more than a couple of
authorities in front of them, they just tum off. High Court judges know the
law usually, anyway, and it is a case of teaching your grandmother to suck
eggs. It is occasionally a bit insulting to High Court judges. (l00)
41 See Campbell, 1976 for a view of provincial solicitors.
42 Morison & Leith, 1992.
[200 ]
Clearly, the picture being painted here is different from that presented in most
AI descriptions of the proposed locations of their AI and law programs. Here the
picture does not accord with the more academic view of the importance of research
and precedent. Law is important, but in a much more routinised manner and all
barristers in this study suggested that their client's case would not be served by too
much emphasis upon legal argumentation.
The consequent of this is, of course, that if legal research is not of primary
importance to barristers, it can only be because the judges do not see it as important.
And if they do not see it in this light, then it is unlikely that they will have a need -
as users - for a rule- or case-based decision support system.
The general tenor of findings in The Barristers' World has been accepted by bar-
risters, and I believe it expresses the situation which the vast majority of barristers
find: they may well wish to actually present legal arguments in court, but the judges
are rarely prepared to accept them. The aspect of the legal process with which
the barrister becomes most involved (and hence, the judge too is most involved
with) is the negotiation over fact. Just why most judges are not predisposed to legal
persuasion is not of primary interest here: it is sufficient to conclude that it is not
likely that they will feel a need for a decision support system which is grounded in
legal rules or cases.
6. Conclusion: What's Wrong with Traditional Decision Support

Techniques?
The presumption of AI proponents is that the computer offers a better way of
doing things. We can see this from the very early days of expert systems: MYCIN
had the worthy aim of attempting to dissuade doctors from prescribing incorrect
antibiotics, a practice which has been leading to the lessening of effectiveness of
these antibiotics against infection. And in legal AI, too, the general assumption is
that the computer will help users to better understand and practice law or, in the
case of sentencing, to sentence in a more rational manner.
This is a large presumption and seems to have developed without any proper
analysis of exactly what is wrong with the more traditional techniques of 'decision
support' - for example, training, education or the use of books. Training and educa-
tion of sentencers has certainly been considered by many43 to be an effective way of
allowing judges to communicate with each other about their sentencing practices.
There seems to be no real criticism of this method, and we can conclude that it is
not unhelpful. And books - particularly textbooks - are used by most lawyers to
provide a grounding in particular points of law, and are well regarded in practice.
Yet, what is it about books which proponents of AI see as such a failing, that they
feel a decision support system which might have only 10,000 words is qualitatively
better than a textbook which contains 150,000 words? Why is less better?
43 For example, the Judicial Studies Board in England and Wales.
[201]
It seems to me that this lack of investigation of the failures of traditional decision

support methods is just another example of the failure to investigate the social
context of the fields in which AI programs are supposed to be effective.
Some 400 years ago we saw a similar situation to the one we have seen over
the past two or three decades. 44 Then, it was the introduction of print technology
which led a group of advocates of the new technologies to advocate new ways of
handling information: in much the same way that today's AI researchers believe
that the new technology arising from the computational metaphor will radically
alter and improve the handling of information and knowledge, those early pioneers
worked with a spirit of hubris attacking the traditional disciplines with relish and
restating their knowledge in the form of rarnist logic. The attempts - as we now
know - were unsuccessful and ramist logic diffused into the intellectual ether and
is now considered a deviant period in the history of ideas. I can't help but feel that a
similar fate awaits today's AI. Certainly, unless it learns to understand what judges
do and what they need, the attempt to aid them will be unsuccessful.
Appendix: What do lye Mean by 'Successful Implementation'?

Discussions about the success of AI usually involve citation of hearsay evidence
concerning some program that someone heard about; was mentioned as an aside
during a conference presentation; has been advertised; or is described in a journal
article by the author of the system. Unfortunately, hearsay evidence is not really
good enough either in court or in science, and means that many of the claims made
for AI are unprovable without further investigation.
It may be that the criticisms in this paper are invalid and that - around the world
- there are a large number of legally based decision support systems in use. If
so, there should be little problem in enumerating these in detail, describing their
usage statistics, length of service, user views, the novelty of technique etc. If there
continues to be little detailed consideration, the sceptic is justified in suggesting
that 'successful implementation' is more difficult than AI proponents suggest.
There are a number of good reasons why the evidence which is provided in a
hearsay context is not useful in a proper consideration of 'successful implementa-
tion'. We can suggest some of the reasons under the following rubrics:
1. NOT EVERY ARTICLE PURCHASED IS USED (THE 'GADGET' CRITICISM)

It is well known that much purchased software is never fully used. An office may
have AI software packages which were bought - perhaps for evaluation - but were
not found to be suitable for their intended task. This is particularly the case where
software which lies towards the user's fringe needs is concerned: an advert may
have encouraged purchase, or indeed the software may have been available at no
cost, and the user has simply acquired a copy, tested it and found it wanting, for one
44 See Chapter 4, Leith, 1990 on Pierre Ramus and his logic.
[202 ]
of many reasons. In this view software is often like the kitchen gadget, purchased
but left in the kitchen cupboard.
2. CLAIMS MADE IN ADVERTS ARE NOT ALWAYS TRUE (THE 'SEXY CLAIM'
CRITICISM)
There are fashions in software just as in clothes. Marketing people are not slow to
link a product with a fashion if they believe that it will boost sales. For example,
the notion of air conditioning being operated according to 'fuzzy logic' principles
is one which has recently been met: just what air conditioning has to do with the
law of the excluded middle is not quite clear, and no doubt the marketing literature
glosses over this. The claim, I suspect, is being made without much linkage to
reality. There are also many claims of this sort found in the selling of software.
We should be wary of software which is advertised in a certain way (e.g., "intel-
ligent") and should - as researchers - attempt to validate the claims made against
the actual operation of the software. Thus: is an advertised 'expert system' really
what we would call an expert system?
3. PROGRAM WRITERS DON'T ALWAYS RETURN TO CORRECT OPTIMISTIC

CLAIMS (THE 'FORGET THE FAILURE' CRITICISM)
The journals of AI and computer science are full of suggestions that programs
constructed by the author of the article are being used. There is usually very little
discussion of who is using this, or how often, etc. I have tried on several occasions
to follow up some of these programs (particularly the ones giving advice in legal
areas) but have found it extremely hard to locate users. Usually it is impossible to
do so. Recourse to the program author provides a later and more measured analysis
describing how the program did not really enter full use. Authors do not hide this
information, but one must approach them to discover it: they rarely follow up their
original article with an article about the program's failure. This is not surprising:
authors are interested in improving their work and want to get onto the next step.
However, readers who take the original article as entirely factual are being led to
incorrect conclusions.
4. THE DIFFERENCE BETWEEN A TRADITIONAL PROGRAM AND AN AI

PROGRAM IS NOT ALWAYS OBVIOUS (THE 'IT WOULD HAVE WORKED
ANYWAY' CRITICISM)
It is welcome to see a program which is successful, but even here we must adopt a
cautious and scientific approach: programs for which substantial claims are being
made require substantial evidence to support these claims. For example, how do we
know that the program only operated through the AI techniques being described?
Programs are difficult objects to analyse, and there is an obvious need for program
[203 ]
authors to take particular care about describing exactly how their program operates.
Some, of course, in legal AI have done this with absolute integrity, but it would be
difficult to believe that all AI researchers routinely did.
5. NOT EVERY COMPANY MAKES A PROFIT OUT OF EVERY ITEM (THE

'BECAUSE SOMEONE SOMEWHERE SELLS IT DOESN'T MEAN IT WORKS'
CRITICISM)
Companies which deal with AI software do not always need to make a profit from
all or some software. It may be that they are simply small part-time businesses
run on a shoe-string selling software which is evaluated by users but never fully
utilised (see the 'gadget' criticism above). Or they may see the software as fash-
ionable and a means of attracting customers (the 'sexy claim' criticism) to other
non-AI software. Because a business exists, does not mean that all its products are
successful.
References
Adams, E. (1972) Courts and Computers, American Judicature Society, Chicago.
Baade, H.W. (ed.) (1991) lurimetrics, Basic Books, New York.
Bainbridge, D.1. (1991) CASE: Computer Assisted Sentencing in Magistrates' Courts, at BILETA
Conference, Warwick.
Buchanan, B.G. and Feigenbaum, E.A. (1981) 'Dendral and meta-dendral: their application dimen-
sion', in Webber, 1981.
Buchanan, B.O. (\986) 'Expert systems: working systems and the research literature', Expert
Systems 3.
Campbell, C. (1976) 'Lawyers and their public', in MacCormick, N. (ed.), Lawyers in their Social
Setting, W. Green, Edinburgh.
Carlen, P. (1976) Magistrates' Justice, Martin Robertson, London.
Cramer, J.A. (1981) Courts and Judges, Sage, London.
Davis, R. (1984) 'Amplifying expertise with expert systems', in Winston, P.H. and Prendergast,
K.A. (eds.), The AI Business: Commercial Uses of Artificial Intelligence, MIT Press, Cambridge,
Mass.
Deedman, C. and Smith, lC. (1991) 'The nervous shock advisor: A legal expert system in case-based
law', in Operational Expert Systems in Canada, Persimmon Press, Elmsford, New York.
Duda, R.O., Hart, P.E., and Reboh, R. (1985) Letter to the Editor, Artificial Intelligence 26(3).
Gordon, T.P. (1995) The Pleadings Game. An Artificial Intelligence Model of Procedural Justice,
Kluwer, Dordrecht.
Hutton, N, Tata, c., and Wilson, J.N. (1995) 'Sentencing and information technology: Incidental
reform?, Intern. J. of Law and Information Technology 2(3)
Lawrence, J. (1994) Sentencing Processes and Decisions: Influences and Interpretative Procedures,
presented at 2nd Annual Conference, Institute for the Study of the Legal Profession, Sheffield.
Leith, P. (1986a) 'Fundamental flaws in legal logic programming', in The Computer 10urnaI29(6).
Leith, P. (1986b) 'Legal expert systems: Misunderstanding the legal process', Computers and Law
(49).
Leith, P. (1990) Formalism in AI and Computer Science, Ellis Horwood, Chichester.
Leith, P. and Hoey, M. (1998) The Computerised Lawyer, Springer Verlag, London.
Leith, P. (1998) Harmonisation of Intellectual Property in Europe, Sweet and Maxwell, London.
[204 ]
Morison, J. and Leith, P. (1992) The Barrister's World and the Nature of Law, Open University Press,
Buckingham.
Munro, C. and Wasik, M. (1992) Sentencing, Judicial Discretion and Training, Sweet & Maxwell,
London.
Newell, A. and Simon, H.A. (1972) Human Problem Solving, Prentice-Hall, Englewood Cliffs, NJ.
Ostberg, 0., Whitakeer, R., and Amick III, B. (1988) The Automated Expert: Technical, Human and
Organizational Considerations in Expert Systems Applications, Teldok, Sweden.
Prakken, H. (1997) Logical Tools for Modelling Legal Argument, Kluwer, Dordrecht.
Prakken, H. and Sartor, G. (1996) 'Rules about rules: Assessing conflicting arguments in legal
reasoning', Artificial Intelligence and Law 4, 331-38.
Rock, P. (1993) The Social World of an English Crown Court, Clarendon, Oxford.
Sartor, G. (1997) 'Logic and argumentation in legal reasoning, Current Legal Theory 25-63.
Sheskin, A. and Grau, C. (1981) 'Judicial response to technocratic reform', in Cramer, 1981.
Shortliffe, E.H. (1981) 'Consultation systems for physicians: The role of artificial intelligence
techniques' , in Webber, 1981.
Shortliffe, E.H. (1992) 'The adolescence of AI in Medicine: Will the field come of age in the '90s?',
in Artificial Intelligence in Medicine 5, 93-106.
Shortliffe, E.H. (1994) 'Health care professional workstations: Where are we now? ... Where should
we be tomorrow?', International Journal of Bio-Medical Computing 34(1-4), 45-55.
Shortliffe, E.H. (1995) 'When decision support doesn't support', in Medical Decision Making 15(2),
187-188.
Simon, E. and Gaes, G. (1989) 'ASSYST - Computer Support Guideline Sentencing', in Proc. 2nd
Int. Con! on AI and Law, ACM Press.
Stamper, R. (1980) 'LEGOL: Modelling rules by computer", in Niblett, B. (ed.), Computer Science
and Law, Cambridge.
Tonry, M. (1992) 'Judges and sentencing policy - the American experience', in Munro & Wasik,
1992.
Turner, A.J. (1992) 'Sentencing in the magistrates' court', in Munro & Wasik, 1992.
Webber, B.L., and Nilsson, N.J. (eds.) (1981) Readings in Artificial Intelligence, Tioga Publishing,
Palo Alto.
Wice, P. (1981) ,Judicial socialization: The Philadelphia experience', in Cramer, 1981.
Woolf, Lord (1996) Final Report: Access to Justice: Final Report to the Lord Chancellor on the Civil
Justice System in England and Wales, copy at www.law.warwick.ac.uklwoo/flreport.
[205 ]
~, ArtificialIntelligence and Law 6: 311-324, 1998.
, 311
Judicial Decisions and Artificial Intelligence
MICHELE TARUFFO
Istituto di Diritto Privato e Processuale, Universitd degli Studi di Pavia, Strada Nuova 65,
27100 Pavia, Italy
1. Some Remarks About Judicial Decision-Making

In order to start a discussion concerning the possibility of applying AI models
to judicial decisions, a necessary preliminary step is to refer to some theoretical
frameworks concerning the practice of judicial decision-making and the reasonings
made by judges. Of course this is not a matter of free choice or of individual pref-
erences: what is needed is to determine which theoretical approach may produce
the relatively best approximation to the empirical phenomenon of decisions made
in the context of the administration of justice. This is an extremely complex and
difficult problem. On the one hand, the experience of the various judicial systems
(and even within a given single system) shows that judicial decision-making in-
cludes an almost infinite range of variations. The factors influencing the ways
in which judicial decisions are made are numerous and include for instance the
format and size of the court (single judge or panels, and so on), the composition
of the court (professional and/or lay judges), several procedural rules, the factual
circumstances of cases, the form and content of the substantive rules governing the
case, the evidence available and the methods and standards used to decide on facts
according to the proofs and to solve legal issues according to the relevant rules and
principles.
On the other hand, the recent history of the western legal culture has produced
a number of different theories concerning judicial decisions. These theories cannot
be fully considered here (for a recent account see for instance Larenz & Canaris
(1995, 99, 133, 187). However, the problem exists of determining which theory
fits best with the experience of judicial decision-making. Of course the answer to
this problem depends also upon the way in which such an experience is described
and rationalized. In a sense, the theories of judicial decisions are rationalizations of
the experience of decision-making. Their number and variety is the consequence
of the variety of concrete experiences and of the variety of philosophical and legal
approaches to the subject matter of judicial decisions. However, this complexity
should not prevent a reasonable choice among the theoretical models of such deci-
sions, in order to assume - as a starting point for further analysis - a model that may
[207 ]
312 MICHELE TARUFFO
be considered as the relatively best approximation to the reality of decision-making

in modem legal systems.
(a) There is no need, however, to choose a whole theory of judicial decision
among the pre-existing and "ready made" theories and to assume it once forever as
valid in all its details. Rather, it may be useful just to stress some basic points
drawing them from the best theories available. One of these points is that the
judicial decision can and should be considered as a set of choices among alter-
native hypotheses of possible decisions (see Roedig, 1969; Taruffo, 1975). The
basic idea is that from the very beginning of a judicial process, and then all along
its development, several possible "projects" or "drafts" of decision are submitted
to the court by the parties, and some other may be built up by the court itself.
Every party alleges her own "version" of the case (or even more than one version)
and - correspondingly - she proposes a hypothesis (and sometimes more than
one hypothesis) for the final decision of the case. When it is vested with such a
power, the court may alternatively find out a further hypothesis of decision that
is different from the ones that are suggested by the parties. Such a "third way of
decision" may then be adopted by the court in order to achieve its final judgment.
Generally, therefore, the situation of the court at the moment of the final decision-
making is characterized by the existence of several (at least two) possible projects
of decision, and by the obligation of the court to choose one of them as the best
possible decision in that context (see, with specific reference to the judgment about
the facts in issue, Taruffo, 1992,266; Taruffo, 1995).
(b) The choice of a hypothesis of decision is far from being a simple problem
because of several reasons. One of these reasons is that a hypothesis of decision is,
in fact, a complex set of statements, each statement being a possible answer to a
relevant issue or law or of fact. A "case" can be imagined as a cluster of issues that
are raised by the parties and possibly also by the court. Such issues may deal with
matters of law and/or matters of fact. Every issue may have two or more possible
answers. Every combination of all the possible answers to the issues determining
a case is a possible "global" hypothesis for the decision of that case. The number
of such global hypotheses is theoretically unlimited. Actually only some of them
become relevant, and it happens when the parties or the court refer to some specific
combinations of possible answers to the several issues of law and fact. In a sense,
the group of the hypotheses that are taken into consideration in order to decide a
case is a very limited subset of the theoretically possible decisions of that case.
Of course such global hypotheses are different not only when they include sets
of totally different answers to every legal or factual issue: in order to be different
two hypotheses may include even different answers to (at least) one relevant issue
(while the answers to the other relevant issues may be the same). However, even
two partially different projects of decision form an alternative: correspondingly the
court has to make a choice between them.
A very general distinction which may be useful to understand the complexity of
hypothetical global decisions is that between legal and factual issues. Legal issues
[208 ]
JUDICIAL DECISIONS AND ARTIFICIAL INTELLIGENCE 313
are those dealing with the choice of the legal rule(s) governing the case and with
the interpretation and application of such rule or rules. Such issues include for
instance the reference to written legal provisions, the use of precedents, the use of
legal canons, the use of legal arguments, and so forth. In a sense, determining the
possible solutions of a legal issue requires dealing with a complex group of related
legal questions and to find out the possible answers to such questions. The set of
such questions and answers is the "legal context" within which the final decision of
the relevant legal matters will be found (see generally Summers & Taruffo, 1991).
Factual issues concern the reliability (i.e., the truth or falsehood) of the statements
about the material facts of the case. The questions of fact are solved on the basis
of the evidence presented, of the proofs emerging from such evidence and of the
inferences connecting evidence and factual statements. Assuming that the "issues
of fact" are a set of factual statements, each possibly being true or false, one may
say that the set of such statements and of their possible values of truth is the "factual
context" within which the final decision concerning the facts of the case will be
chosen (see Taruffo, 1992,217,293; Taruffo, 1995).
(c) The judge's reasoning concerning the various hypotheses about the legal
and factual issues has a basically dialogical structure. In fact it is based upon the
comparison, the contrast and a final choice among the competing solutions of such
issues. There are three major explanations for this dialectical/dialogical structure.
An explanation is that most of the hypotheses of decision are proposed by the
parties in the course of the proceedings preparing the final judgment. The process
may be interpreted as a contest or a conflict between the (two or more) parties: from
the point of view of what the parties "say" it is a dialogue. Each party sets forth
her own "theory (or theories) of the case" that is made of the groups of statements
of law and of fact that are proposed as a valid and reliable hypothesis of decision.
Each "theory of the case" often includes also the objections and rebuttals directed
against the "theory of the case" proposed by the other party (see Taruffo 1994, 389;
Taruffo, 1995, 789). At least in principle, each affirmative or negative statement
made by a party is supported by arguments. Many of these statements are in fact
answers and objections to the statements made or to the issues raised by an adverse
party. The dialectical relationship between the parties is carried on during the whole
development of the judicial process. Such an interaction of the parties is the main
source of the various hypotheses for the final decision, among which the court will
make its choice.
The second explanation of the dialogical structure of the court's reasoning is
that the set of hypotheses among which the court has to choose is deeply influenced
by the fact that the decision will have effects "between the parties", and that every
version of the case proposed by a party is by definition in conflict with the versions
proposed by the other party or parties. In a sense, in order to choose a decision
the court has "to run again through" the dialogue that has been carried on by the
parties. The court shall compare arguments and counterarguments, statements and
[209 ]
314 MICHELE TARUFFO
objections, affirmations and negations, in order to select the relatively most reliable
answers to the legal and factual issues of the case.
The third explanation is that the decision-making reasoning may be interpreted
as a dialogue of the judge with herself. It is a widespread commonplace that de-
ciding a case is a complex play of "trial and error". There is some truth in this
commonplace, but one should stress that the judge plays this game with herself:
the judge starts from a hypothesis (either by deriving it by a party's argument or
stating it on her own motion), and then she "tries" this hypothesis in order to check
it and to decide whether it is wrong or correct. The "hermeneutic circle" (see e.g.,
Larenz & Canaris, 1995,288,298) involved in every procedure of decision-making
has a basically dialogical structure. One may think of a dialogue of the interpreter
with the interpreted text, when textual interpretation is involved, or more generally
of a dialogue of the decision maker with herself when the problem is of checking
the correctness of a hypothesis for a possible decision.
2. Some Remarks About the Justification of Judicial Decisions

After having made the choices that are needed to achieve the final decision, the
judge is usually under the obligation to justify such a decision. In all the developed
judicial systems the judge is expected to deliver an opinion (usually in a written
form, sometimes orally) in which she expresses the arguments supporting her final
judgment (see generally Taruffo, 1975,319).
(a) The justificatory opinion delivered by the judge is supposed to be a "justifi-
cation" in a proper sense, that is an articulated reasoning expressing the arguments
on the basis of which the judgment should appear as valid, reliable, just and rea-
sonable. In a word, the opinion should show that the decision is "reasoned", i.e.,
well grounded upon "good reasons".
The complexity of the proceeding of decision-making influences the structure
and the content of the opinion justifying the decision. This is not to say that there
is any kind of direct correspondence or identity between the reasoning of deci-
sion making and the justificatory reasoning expressed by the judge in her opinion.
On the contrary, it is commonly said that these two reasonings of the judge are
structurally and functionally different: decision-making is a sort of "context of
discovery" aimed at finding out (i.e., at choosing among alternatives) the correct
decision by way of trial and error, hypothesis and control, and so forth, while the
justificatory opinion is a sort of "context of justification". It assumes the decision as
a starting point, and it is aimed at showing that the decision is "good" on the basis
of a set of reasoned justificatory arguments (Taruffo, 1975, 118). However, when a
decision is complex, because in order to achieve the final judgment the judge had
to make a complicated series of choices among various alternatives concerning
several issues of law and of fact, the consequence is that the justification of such
a decision is also complex. The judge's opinion shall be made of a complex set
of arguments justifying the choices made by the judge about any relevant issue of
[210 ]
law and of fact. If the decision may be imagined as a set of statements, each of
them expressing the solution of a relevant issue that has been chosen by the judge,
then the justificatory opinion may be imagined as a set of arguments in which one
or more arguments deal with each of those statements. A complete justification
is the one in which every relevant statement included in the decision is expressly
and properly justified. A consistent justification is the one in which the arguments
used do not conflict with each other. A sufficient or adequate justification is the
one in which every statement is well grounded upon good justificatory arguments.
A coherent justification is the one in which the arguments used fit well with the
nature of the issues decided. Thus, the statements expressing the interpretation of
the governing legal rules that has been chosen by the judge as a proper and correct
interpretation in that case should be justified on the basis of appropriate legal ar-
guments grounded upon relevant legal materials (rules, precedents, canons, and so
forth). Correspondingly, the statements expressing the version of the material facts
that have been found to be "true" by the judge should be justified on the basis of
the relevant evidence, of the standards used to assess the value and credibility of
proofs, and of the inferences supporting the final choice of the judge (see generally
Taruffo, 1975, 265, 430, 548).
The actual practices concerning justificatory opinions are rather different in the
various legal systems. There are in fact several styles of judicial opinions accord-
ing to different traditions and legal cultures (see Summers & Taruffo, 1991, 461).
However, the function of such opinions is roughly the same in every legal system.
Correspondingly, the basic structure of such opinions tends to be the same, when
one looks at the most important features of the judge's reasoning. At least, a general
rational model of justificatory opinion may be imagined by combining the analysis
of the structure of the decision and the consideration of the purpose and function
of the justification of a judgment.
(b) A very important feature of many judicial opinions is that they have a dia-
logical structure. Such a feature is more or less evident in the various systems:
for instance, it is less visible in French judgments and more visible in Italian
judgments, but it exists, at least at a deep level, in every case. Such a dialogical
structure of justificatory opinions derives from two main factors that are in part
overlapping upon each other.
One of these factors is that the justification of a judgment is - among many other
things - also the answer that the judge gives to the parties of the case. This factor
is specially clear when the parties are required to set forth their own version of the
case in the form of specific "grounds" or issues, because the judge will expressly
consider each issue and the arguments proposed by the parties, and he will answer
to such arguments by accepting or rejecting them and stating the reasons why he
accepts or rejects these arguments. In a sense, therefore, the opinion is a sort of
dialogue between the judge and the parties, in which the judge deals with the
questions put by the parties and considers the arguments used by the parties to
[211 ]
316 MICHELE TARUFFO
support the solutions they propose for these questions. Then she gives an answer
to these questions and explains the reasons supporting her answer.
The second factor favouring the dialogical structure of justificatory opinions
is that the judge is required to set forth the reasons grounding her own choices.
This is to say that the judge will use arguments of several kinds in order to build
up such a justification. The problem is complex, however, because justifying a
choice may require a complicated reasoning. On the one hand, the judge should
use arguments positively supporting the choice (i.e., the statement resulting from
it) which has to be justified. It is necessary in any kind of justification, and it is
specially necessary in judicial opinions. On the other hand, a choice is not com-
pletely and properly justified if the alternatives that have been rejected are not taken
into due consideration. A choice is not "completely" justified if the justification
deals only with the reasons supporting the alternative that has been choosen. The
decision should be justified also on the basis of the reasons according to which
the other relevant alternatives have been rejected (see Prakken & Sartor, 1997,
§2.1). In fact, a rejected alternative could have been better than the alternative that
has been accepted. In order to show that the decision that has been finally chosen
is the relatively best one, the judge should demonstrate that there were no better
alternatives available. Also in the context of justification, therefore, and not only in
the context of decision-making, the comparison and the critical analysis of all the
relevant possibilities is extremely important. This is to say that the judge should
confront herself with the other possible decisions and with the arguments possibly
supporting such decisions, with the aim of arguing and proving that such arguments
were not valid, reliable or persuasive.
3. Judicial Decision.Making and Artificial Intelligence

If one considers the evident features of complexity, variability, flexibility and dis-
cretion that are typical of judicial decisions, any approach aimed at interpreting the
judicial reasoning according to logical rules and models may appear as doomed to
failure. In fact, the history of the logical theories of judicial reasoning is largely
a history of misunderstandings, errors, manipulations and defeats (see e.g., Sartor,
1997). The long history of the unsuccessful but numerous attempts to represent
the judicial decision as a syllogism or as a chain of syllogistic steps is a very well
known example of it. Still at present a large amount of studies is devoted to the
problem of connecting logic and decision-making and of building up appropriate
logical frameworks for judicial reasoning (see again Sartor, 1997). The distrust in
the possibility of a logical formalization of judicial reasoning may be even more
intense when the problem is whether such a reasoning may be interpreted and for-
malized in terms of computerized logic or, more generally, in terms of AI models.
On the one hand, one may observe that the main attempts to "computerize" the
reasoning of the judge were so rough, and unable to interpret the complex nature
of decision-making, that they could not succeed in producing reliable models of
[212 ]
the judge's reasoning (see generally Leith, 1997). These attempts, one might add,
are a good proof of the impossibility of interpreting such a reasoning in terms
of AI. On the other hand, one may consider that the decision-making procedure
is so complex, variable, uncertain, fuzzy and value-laden, that it could never be
reduced to logical models. Any logical model, one might say, would necessarily
leave aside important features of the decision-making reasoning that cannot be
reduced to logical forms. Therefore, such a model would be basically false as a
description and inappropriate as a prescriptive model for judges.
In fact, the idea that the judge's reasoning could or should be completely re-
duced to simple logical models, or to a narrow group of computerized calculi, is
clearly untenable. However, the problem is not to find out a ready-made and simple
model but to verify whether AI is - or may be - able to interpret the reasoning used
by judges or at least some features of such a reasoning. It is rather clear that this
cannot be made by means of simple and easy logical models. On the other hand,
research in AI is rapidly evolving and already offers a wide and growing inventory
of analytical tools: some of them are hopefully fit with the problems of judicial
reasoning. Therefore, at present the problem is neither of making a priori acts of
faith in the all-encompassing virtues of AI nor of making a priori acts of disbelief
in any possibility of using AI as a means to elicit, interpret and represent the judge's
reasoning. Both attitudes, in fact, would be misplaced and unjustified on the basis
of the current "state of the art" of AI applied to the problem of judicial decision-
making. Rather, one should take into account the emerging trends in this area and
consider whether they appear well oriented and possibly fruitful.
It is impossible to make here a detailed analysis of the research made in the last
years in the field of AI with reference to several aspects of the judge's reasoning.
Moreover, this research is growing and developing almost day by day, so that such
an attempt would be rapidly outdated. It may be useful, however, to look in syn-
thetic and general terms at the main directions of this research, in order to sketch
some general remarks and evaluations.
(a) A relevant group of studies is aimed at standardizing procedures used by
courts according to computerized models. The goal that is pursued by this re-
search is to facilitate the management of some procedures by the courts by using
computers as means to implement such procedures quickly and economically. The
basic idea is of having a program corresponding with the relevant features of the
procedures involved, so that a procedure may be created and managed by the com-
puter according with this program. In each single case one should only insert the
individual data, and the outcome - that is: an order, an act, a decree - should be
automatically produced. In some cases some interesting results have been achieved
(see Branting, Lester & Callaway, 1997). However, the experience made so far
shows that there are problems and limits to deal with in order to develop this
perspective of application of AI to the functioning of judicial proceedings.
An important factor, and a significant limit, is that the procedures involved
should be simple and able to be repeated in many cases without significant varia-
[2l3 ]
318 MICHELE TARUFFO
tions. This is an important condition for the manageability of the program. A simple
procedure may be more easily standardized in the form of a computer program for
the trivial reason that each step of the procedure has to be translated into a part of
the program. The lower the number of the procedural steps considered, the easier is
to put them into a software. A further important feature of simplicity is the limited
number of alternatives that are possible or should be admitted at any point of the
procedure in which a choice has to be made. If at a given point a procedure may be
carried on following different paths, one must know in advance which paths may
be followed in order to insert them in the program. If such alternatives are two or
three rather than some dozens it is much better. It is even better if in a procedure
there are few "points of choice" rather than dozens of them. At any rate, each
point of choice and all the alternatives admitted for each choice must be defined
in advance in order to have a complete formalization of the procedure considered.
Some choices may be left "open" (i.e., some alternatives not previously determined
may be admitted), but every "openness" entails a lack of efficiency of the system
because it requires an ad hoc consideration by the user.
The procedure should also be able to apply without significant variations in a
relatively high number of cases. If a procedure tends to vary in the concrete cases to
which it applies it is difficult or impossible to formalize it in a complete way or to
adapt it to the different specific situations. Then the procedure should apply to areas
in which concrete cases tend to occur substantially in the same relevant terms. On
the other hand, such cases should be rather numerous if the computerized procedure
has to be an efficient and economic way to deal with the procedures involved.
These factors of simplicity, repetitiveness and frequency of procedures in like or
identical cases are rather easy to find in several areas of bureaucratic administra-
tions, in private organizations as well as in public agencies. Such situations may be
found also in the field of judicial procedures, but here some further problems arise.
On the one hand, simple, repetitive and frequent cases exist, but - unfortunately -
they are neither the most frequent nor the "normal" situation. The so-called "easy
cases" often are not easy enough to be standardized in terms of computerized
models. A fortiori such a standardization is extremely difficult or impossible to
achieve in the "average" judicial case, let alone in hard cases. It seems, therefore,
that the type of computerization we are considering may be useful but only in a
relative narrow area of judicial practice. On the other hand, one must consider that
the simplicity of a procedure is not given a priori, since it derives from the legal
regulation of this procedure. Then the simplicity or complexity of a procedure is
determined by the authority that is vested with the power to regulate it. Therefore,
in a sense the possibility of translating a procedure into a software program depends
upon the choice made by the lawgiver about the type of procedure applied in a given
situation. For instance, collection of debts may be made by means of a very simple
procedure or a complex one depending upon the lawgiver's choice as to whether it
is better to facilitate the satisfaction of the creditor rather than the debtor's defence.
[214 ]
Moreover, the simplicity of the procedure is a matter of degree and of evalua-

tive choices. A "matter of degree" means that there are no absolute and a priori
standards of simplicity or complexity concerning procedures. Procedures are more
or less simple or complex according to a wide number of factors, including the
choices of the lawgiver, the behaviours of the subjects involved, the variety of
cases, and so forth. Then the level at which AI methods may be applied cannot
be determined in advance. Simplicity or complexity of a procedure is a matter of
evaluation in several ways. An important evaluative dimension is whether a proce-
dure fits adequately with the nature of cases to which applies. A simple procedure
may be inappropriate for complex cases, unless it is flexible enough to be adapted
case by case to changing circumstances and to difficult issues. On the other hand, a
complex procedure may be inappropriate for simple and repetitive cases because it
would be exceedingly burdensome and expensive. A major danger existing in the
trend to simplify judicial procedures (eventually in order to standardize them and
to translate them into software programs) is, however, that of "losing" the relevant
complexity of actual cases or of having procedures that, not being able to grasp and
to deal with such a complexity, would lead to inaccurate or inappropriate decisions.
Simplicity, in fact, is not a value by itself.
If one keeps these remarks in his mind while thinking of judicial procedures, she
may easily come to the conclusion that the standardization of procedures in terms
of software programs may be applied in the field of judicial proceedings, but it
can cover only some narrow areas of judicial practice. Roughly speaking, these are
the areas in which the administration of justice is more similar to the bureaucratic
administration with regard to the procedures employed and to the repetitiveness of
concrete cases. But most civil or criminal cases cannot be reduced to the ideal type
of a simple case deserving to be tried and decided by a simple and standardized
procedure. Average and complex cases usually require complex procedures, but
such procedures in their tum require - in order to be formalized and standardized -
such a high level of sophistication in the methods of AI that at present they cannot
even be imagined.
(b) Among the significant instances of application of the methods of artificial
intelligence there is the attempt to rationalize discretionary decisions. The main
field of application of such a rationalization has been the problem of sentencing in
criminal cases (see Schild, 1997; Tata, 1997).
The attempt to rationalize the discretionary judgments of courts raises several
important issues. One should distinguish, for instance, between strong discretion
and weak or regulated discretion. Strong discretion exists when the judge is com-
pletely free to choose her own decision within a theoretically unlimited range of
alternatives, and to take into account only the peculiar features of the individ-
ual case. Weak discretion exists when the judge is relatively free to choose her
own decision but such a choice should be made either within a previously deter-
mined inventory of alternatives, or within a range of quantitative possibilities with
pre-determined minimum and maximum, or even when the judge should make a
[215 ]
320 MICHELE TARUFFO
discretionary decision according to standards or principles governing the matter.

One may speak also of regulated discretion when these limits, rules or standards
are provided for and imposed by the law (about discretion see generally Christie,
1987).
There is no need to discuss analytically such a distinction here, nor is it neces-
sary to examine the various types of weak discretion. However, distinctions con-
cerning different kinds of discretion should be taken into account while discussing
methods to rationalize discretionary judgments.
It should be considered that strong discretion cannot be rationalized by means
of any logical tool or framework. The basic feature of such a discretion, in fact, is
just that the judge is vested with the power to decide each individual case just by
paying attention to any relevant feature of such a case and by choosing the premises
or the standards for the decision with a complete freedom of choice. The judge
shall determine case by case which principle, standard, criterion or canon is more
appropriate as a basis for such a specific decision. It seems clear that this choice
cannot be rationalized ex ante, for instance by imposing rules and procedures of
choice to the decision-maker, without destroying it or transforming it into anything
else, that is into some form of weak or regulated discretion. At most, the judge may
be required to justify ex post her choices. The judge may make an ex post justifi-
cation of a decision based on strong discretion by stating expressly the standards
and criteria she used in order to reach her decision, and by referring to the features
of the specific case that she has considered relevant. The judge may show that,
given some relevant circumstances of the case and some standards of evaluation,
the decision she has made is reasonable and valid according with its premises. The
judge might even deliver an external justification of these premises by referring
them to higher canons of choice (such as moral or political values justifying the
assumption of a principle governing the use of discretion in a particular case),
and by giving reasons justifying why she considered some features of the case
(for instance: the gender or the social condition of a subject) as relevant for her
decision. However, this is not an a priori rationalization of the decision-making
process (which is probably impossible in terms of logical rules and models) but
only an a posteriori rationalization of a decision in terms of arguments that seem
appropriate to present the decision as founded upon "good reasons". To the extent
that such a rationalization can be made by using methods of AI, it belongs to the
area of the analysis of the modes of reasoning by arguments (see infra, (c)), rather
than to the area of the rationalization of discretionary evaluations.
When weak or regulated discretion is involved, it seems that there is more room
for an ex ante rationalization of discretionary choices. In fact, in such cases there
are pre-existing standards and criteria of decision which determine ex ante the
borders of judicial discretion and some rules or guidelines governing the choices
of the judge. In other terms, there is a set of prescriptions to follow, or a set of
alternatives to choose, that define the context in which the discretionary decision
shall be made. It is just the case of criminal sentencing when the law determines,
[216 ]
for instance, a maximum and a minimum of imprisonment for each type or crime,
or it prescribes that the judge should refer to some factors (such as age, social
and mental conditions, prior convictions, attitude to commit future crimes, and so
forth), while determining the size of the criminal sanction in a given case. Some-
times a "tariff" is used as a means to regulate the court's discretion in sentencing
(see Schild, 1997, §6.2.1.). The usual standards for sentencing can be even more
detailed and better defined if all the relevant standards are combined in a precise,
complex and sophisticated set of analytical rules producing more detailed classifi-
cations, according to which every case may find its own appropriate pigeonhole in
which the "proper" sentencing may be determined.
Building up complicated combinations of standards and principles, and analyt-
ical classifications aimed at defining sub-sets of cases (and perhaps several layers
of subsets determined per genus et dijJerentiam), is theoretically possible. Perhaps
it is a functional way to rationalize the decision of the judge at the moment of
sentencing, since the judge is supposed to be able to put the single case into the
proper and exact pigeonhole, in which - so to say - a label saying for instance "6
months" or "20 years" will be found. Perhaps such a method may be useful, if the
purpose is to maximize the rationalization of the judge's discretion in determining
the proper criminal sanction. It may even be unavoidable if the methods of AI have
to be used as means to rationalize the judge's evaluations, at least to the extent that
these methods require a previous precise determination of the alternatives among
which the judge will choose her decision in the individual case. In fact, some
programs aimed at applying AI methods to sentencing seem to have a feature in
common: this feature is that, explicitly or implicitly, by law or by logic, by statute
or by practice, many hypotheses are stated by combination of the relevant standards
of evaluation, and a more or less wide inventory of alternatives is under the judge's
eyes. The main idea is that the judge will choose one of such alternatives, within
a context in which any case should theoretically find its own proper "pigeonhole",
and therefore its own proper sanction, inside a consistent and well-organized set of
situations. Such a classification is the result of a rational and logical combination
of the standards and principles applying to sentencing in that legal system.
However, this system tends to produce extremely complicated combinations
(see Schild, 1997, §3.2.). A further problem is that, in order to be classified or
referred to a typical situation, individual cases should be considered as "similar"
and corresponding to a type (see Tata, 1997). But this does not solve the problem
of judicial discretion, because such a discretion will be used when the judge de-
termines whether a case is similar to another case or whether a case fits with a
type.
We may concede, at any rate, that there may be efficient AI methods to ratio-
nalize the judge's sentencing. However, the danger is of eliminating the judge's
discretion rather than rationalizing it. In fact, AI methods may be efficient insofar
I
as they reduce or eliminate the vagueness, the fuzziness, the, open texture and
the indeterminacy of the standards governing the practice of sentencing. Such a
[217 ]
322 MICHELE TARUFFO
reduction or elimination may even be considered as a positive change in the field

of criminal law, to the extent that it may increase uniformity and foreseeability
(and then certainty and equality) in sentencing, and correspondingly it may de-
crease or eliminate subjectivity, uncertainty, variability and even inequal treatment
in sentencing. From this point of view this may even be considered a reasonable or
rational change, if it is assumed that values such as uniformity, and so forth, should
take the place of case by case evaluations.
The main point is, however, that under the label of rationalizing the use of
discretion in sentencing what is actually done is a chance in criminal policy that is
achieved by reducing or excluding the judge's discretion. Reducing or excluding
discretion in sentencing is not a problem of method, nor is it simply an instance in
the application of AI: it is basically a problem concerning the policy of sanctions
in criminal law. This problem involves a number of complex issues such as: should
general standards prevail or not upon the consideration of the peculiar features of
the person involved? to what extent should judges be trusted in using discretion
in a given legal system? which guarantees mayor should be enacted in order to
prevent abuses of judicial discretion? And so forth. Anyone can perceive that these
are extremely important problems the solution of which affects the life and liberty
of people.
(c) A further and rather new field for the application of AI methods is the analy-
sis of judicial reasoning focusing upon the arguments used to make and to justify
decisions. It is not necessary to make here a detailed analysis of the research made
in this area (see for instance, Rage, 1996; Freeman & Fairley, 1996). It is worth
stressing, however, that this research deals with some of the most interesting topics
of the judge's reasoning, such as the use of precedents (see Prakken & Sartor,
1997), the dynamics of interpretive arguments (Sartor, 1994), the use of arguments
and the solution of their conflicts (see Prakken & Sartor, 1996; Kowalski & Toni,
1996; Sartor, 1997, §1O.2), the structure of chains of arguments (see Sartor, 1997,
§9; Sartor, 1994), and so forth. In order to deal with the complexity and the varying
structure of judicial reasoning, such research employs appropriate logical tools
such as non-monotonic logic and special formalizations.
So far this area of research is far from providing with a complete analysis of
judicial reasoning, nor does it include a detailed set of logical frameworks applying
to all the relevant features of judicial decision-making and of the justification of
judicial decisions. There are, however, some relevant reasons of interest in this
dimension of analysis.
First of all, this research finally sets aside a number of logical models that
proved to be unsuccessful and unreliable in the history of the legal culture concern-
ing judicial decision-making. The most important victim is the syllogistic model,
but it should not be regretted: it has always been an unreliable picture of how judges
make and justify their decisions (see e.g., Taruffo, 1975, 126, 149). Moreover, this
research shows that the judge's reasoning cannot be reduced to simple, one-sided,
deductive, repetitive, fixed, pre-determined and all-encompassing logical models
[218 ]
of any kind. On the other hand, they show that judicial reasoning is open to logical
and rational analysis although it cannot be interpreted in terms of elementary logic.
A relevant consequence is that the traditional rigid alternative between deduction
and irrationality is clearly wrong and misleading. A decision can be rational and
logically founded even when it is not (as it usually happens) deductive.
The type of analysis we are considering is based upon the use of several logical
tools. Also because of this, however, it pays due consideration to the complexity
and the heterogeneity of the judge's reasonings. In fact, one of the flaws of the
traditional theories of judicial decision-making is that they presume to interpret
extremely complex phenomena just by using very simple and rough tools (such as
the principles of basic logic or the relics of philosophy that lawyers studied at the
high school). As we stressed above, on the contrary, there are several factors of
high-level complexity both in judicial decision-making and in the justification of
judicial decisions. Such complexities require, in order to be properly understood
and explained, the use of sophisticated logical and analytical tools. In fact, the
modem theory of judicial reasoning becomes more an more complex as it tends to
deepen the study of the actual features of such a reasoning in logical and rational
terms (for a significant example see Peczenik, 1996).
If a judicial decision is conceived as the outcome of a cluster of choices concern-
ing the solution of legal and factual issues, a rational analysis of such a decision
should deal with the logic of rational choices as they are made in the peculiar
context of the administration of justice. Thus, for instance, the rational structure
of reasoning concerning facts, evidence, inferences about factual statements, and
so forth, requires to be analyzed with appropriate conceptual tools. Some of these
tools may be drawn, with some adaptations, from the inventory of the concepts
used to computerize the flow of information or the flux of knowledge (see for
instance Taruffo, 1995,804). Similarly, the complex structure of the reasoning by
which the judge makes his choices concerning the determination, the interpretation
and the application of the legal rules governing the case may be studied by means
of the logic of rational argumentation and of the logic of deontic or prescriptive
statements.
More specifically, it seems that the models of non-monotonic logic and the
analysis of the defeasibility of arguments fits very well with the dialogical structure
of the judge's reasoning that has been stressed above (see supra, §lc and 2b).
Advancing arguments and counter-arguments is one of the fundamental factors of
judicial proceedings and, correspondingly, of judicial decisions as well. Therefore,
understanding the dynamics of such proceedings and decisions requires the use of
appropriate logical tools. The study of the modes of arguments and of their dialec-
tical conflicts seems to be specially promising both for realizing how decisions are
actually achieved and justified and for extending the use of AI methods in the area
of judicial reasoning (see specially Prakken & Sartor, 1997, §4, 6; Sartor, 1997, §8,
11; Prakken & Sartor, 1996).
[219 ]
324 MICHELE TARUFFO
As we have stressed above, this research is still at its beginning and the purpose
of building up a complete AI theory of judicial decision-making is very far from
being achieved. However, it seems clear that the use of diversified and flexible
logical models is a promising way of dealing with the practice of making and
justifying rational decision in the judicial context.
References
Branting, L.K, Lester, J.e., and Callaway, Ch.B. (1997) Automatic Judicial Document Drafting: A
Unification-Based Approach. In this issue, pp 111-149.
Christie, G.e. (1986) 'An essay on discretion', Duke L.J. 747.
Freeman, K, and Farley A.M. (1996) 'A model of argumentation and its application to legal
reasoning', AI & Law 4, 163.
Rage, J. (1996) 'A theory onegal reasoning and a logic to match', AI & Law 4, 199.
Kowalski, R.A. and Toni, F.: (1996) 'Abstract Argumentation', AI & Law 4, 275.
Larenz, K and Canaris, W. (1995) Methodenlehre der Rechtswissenschaft. 3.Aufl New York:
Springer.
Leith, Ph. (1997) The Judge and the Computer: How Best "Decision Support"? In this issue, pp.
289-309.
Pecezenik, A. (1996) 'Jumps and logic in the law', AI & Law 4, 297.
Prakken, H. and Sartor, G. (1996) 'A dialectical model of assessing conflicting arguments in legal
reasoning', AI & Law 4,331.
Prakken, R. and Sartor, and 1997. Modelling Reasoning with Precedents in a Formal Dialogue Game.
In this issue, pp. 231-287.
Roedig, J. (1969) Die Denkform der Alternative in Jurisprudenz. Berlin: Springer.
Sartor, G. (1994) 'A formal model of legal argumentation', Ratio Juris 7, 177.
Sartor, G. (1997) 'Logic and argumentation in legal reasoning', Current Legal Theory 25.
Schild, U.J. (1997) Criminal Sentencing and Intelligence Decision Support. In this issue, pp. 151-
202.
Summers, R.S. and Taruffo, M. (1991) In MacCormick, N., and Taruffo, M. (eds.), Interpreting
Statutes: A Comparative Study, p. 461. Rants: Dartmouth.
Taruffo, M. (1975) La motl'vazione della sentenza civile. Padova: Cedam.
Taruffo, M. (1992) La prova dei fatti giuridici. Nozioni generali. Milano: Giuffre.
Taruffo, M. (1994) 'Involvement and detachment in the presentation of evidence', in Krawietz, W.,
MacCormick, N., and von Wright, G.R. (eds.), Prescriptive Formality and Normative Rationality
in Modem Legal Systems. Festschriftfor Robert S. Summers, p. 385. **
Taruffo, M. (1995) 'Elementi per un'analisi del giudizio di fatto', Rivista trimestrale di diritto e
procedura civile 785.
Tata, e. (1977) The Application of Judicial Intelligence and "Rules" to Systems Supporting
Discretionary Judicial Decision-Making. In this issue, pp. 203-230.
[220]
Law and Philosophy Library
1. E. Bulygin, J.-L. Gardies and 1. Niiniluoto (eds.): Man, Law and Modem Forms of Life. With
an Introduction by M.D. Bayles. 1985 ISBN 90-277-1869-5
2. W. Sadurski: Giving Desert Its Due. Social Justice and Legal Theory. 1985
ISBN 90-277-1941-1
3. N. MacCormick and O. Weinberger: An Institutional Theory of Law. New Approaches to Legal
Positivism. 1986 ISBN 90-277-2079-7
4. A. Aamio: The Rational as Reasonable. A Treatise on Legal Justification. 1987
ISBN 90-277-2276-5
5. M.D. Bayles: Principles of Law. A Normative Analysis. 1987
ISBN 90-277-2412-1; Pb: 90-277-2413-X
6. A. Soeteman: Lagic in Law. Remarks on Logic and Rationality in Normative Reasoning,
Especially in Law. 1989 ISBN 0-7923-0042-4
7. C.T. Sistare: Responsibility and Criminal Liability. 1989 ISBN 0-7923-0396-2
8. A. Peczenik: On Law and Reason. 1989 ISBN 0-7923-0444-6
9. W. Sadurski: Moral Pluralism and Legal Neutrality. 1990 ISBN 0-7923-0565-5
10. M.D. Bayles: Procedural Justice. Allocating to Individuals. 1990 ISBN 0-7923-0567-1
11. P. Nerhot (ed.): Law, Interpretation and Reality. Essays in Epistemology, Hermeneutics and
Jurisprudence. 1990 ISBN 0-7923-0593-0
12. A.W. Norrie: Law, Ideology and Punishment. Retrieval and Critique of the Liberal Ideal of
Criminal Justice. 1991 ISBN 0-7923-1013-6
13. P. Nerhot (ed.): Legal Knowledge and Analogy. Fragments of Legal Epistemology, Hermeneu-
tics and Linguistics. 1991 ISBN 0-7923-1065-9
14. O. Weinberger: Law, Institution and Legal Politics. Fundamental Problems of Legal Theory
and Social Philosophy. 1991 ISBN 0-7923-1143-4
15. J. Wroblewski: The Judicial Application ofLaw. Edited by Z. Bankowski and N. MacCormick.
1992 ISBN 0-7923-1569-3
16. T. Wilhelmsson: Critical Studies in Private Law. A Treatise on Need-Rational Principles in
Modem Law. 1992 ISBN 0-7923-1659-2
17. M.D. Bayles: Hart's Legal Philosophy. An Examination. 1992 ISBN 0-7923-1981-8
18. D.W.P. Ruiter: Institutional Legal Facts. Legal Powers and their Effects. 1993
ISBN 0-7923-2441-2
19. J. Schonsheck: On Criminalization. An Essay in the Philosophy of the Criminal Law. 1994
ISBN 0-7923-2663-6
20. R.P. Malloy and J. Evensky (eds.): Adam Smith and the Philosophy of Law and Economics.
1994 ISBN 0-7923-2796-9
21. Z. Bankowski, I. White and U. Hahn (eds.): Informatics and the Foundations of Legal Reason-
ing. 1995 ISBN 0-7923-3455-8
Law and Philosophy Library
22. E. Lagerspetz: The Opposite Mirrors. An Essay on the Conventionalist Theory of Institutions.
1995 ISBN 0-7923-3325-X
23. M. van Hees: Rights and Decisions. Formal Models of Law and Liberalism. 1995
ISBN 0-7923-3754-9
24. B. Anderson: "Discovery" in Legal Decision-Making. 1996 ISBN 0-7923-3981-9
25. S. Urbina: Reason, Democracy, Society. A Study on the Basis of Legal Thinking. 1996
ISBN 0-7923-4262-3
26. E. Attwooll: The Tapestry of the Law. Scotland, Legal Culture and Legal Theory. 1997
ISBN 0-7923-4310-7
27. J.e. Hage: Reasoning with Rules. An Essay on Legal Reasoning and Its Underlying Logic.
1997 ISBN 0-7923-4325-5
28. R.A. Hillman: The Richness of Contract Law. An Analysis and Critique of Contemporary
Theories of Contract Law. 1997 ISBN 0-7923-4336-0; 0-7923-5063-4 (Pb)
29. e. Wellman: An Approach to Rights. Studies in the Philosophy of Law and Morals. 1997
ISBN 0-7923-4467-7
30. B. van Roermund: Law, Narrative and Reality. An Essay in Intercepting Politics. 1997
ISBN 0-7923-4621-1
31. 1. Ward: Kantianism, Postmodernism and Critical Legal Thought. 1997
ISBN 0-7923-4745-5
32. H. Prakken: Logical Tools for Modelling Legal Argument. A Study of Defeasible Reasoning
in Law. 1997 ISBN 0-7923-4776-5
33. T. May: Autonomy, Authority and Moral Responsibility. 1998 ISBN 0-7923-4851-6
34. M. Atienza and J.R. Manero: A Theory of Legal Sentences. 1998 ISBN 0-7923-4856-7
35. E.A. Christodoulidis: Law and Reflexive Politics. 1998 ISBN 0-7923-4954-7
36. L.M.M. Royakkers: Extending Deontic Logic for the Formalisation of Legal Rules. 1998
ISBN 0-7923-4982-2
37. J.J. Moreso: Legal Indeterminacy and Constitutional Interpretation. 1998
ISBN 0-7923-5156-8
Other books on Legal Theory
H. Prakken and G. Sartor (cds.): Logical Models of Legal Argumentation. 1998
ISBN 0-7923-4413-8
G. Postema (ed.): Racism and the Law. The Legacy and Lessons of Plessy. 1998
ISBN 0-7923-4665-3
N. MacCormick (ed.): Constructing Legal Systems. "European Union" in Legal Theory. 1998
ISBN 0-7923-4731-5
KLUWER ACADEMIC PUBLISHERS - DORDRECHT / BOSTON / LONDON

Judicial Applications of Artificial Intelligence: Springer-Science+Business Media, B.V

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Judicial Applications of Artificial Intelligence: Springer-Science+Business Media, B.V

Uploaded by

Copyright:

Available Formats

JUDICIAL APPLICATIONS OF

Reprinted from Artificial Intelligence and Law 6(2-4) 1998

Springer-Science+Business Media, B.V.

ISBN 978-90-481-5136-3 ISBN 978-94-015-9010-5 (eBook)

Printed on acid-free paper

All Rights Reserved

No part of the material protected by this copyright notice may be reproduced or

G. SARTOR and L. KARL BRANTING / Introduction: Judicial Appli-

Introduction: Judicial Applications of Artificial

1. Modeling judicial tasks

No form of legal reasoning seems to depend more heavily on uniquely human

2. The papers in the special issue

analogizing and distinguishing a precedent). Finally, the paper presents a method

Automating Judicial Document Drafting: A

JAMES C. LESTER and CHARLES B. CALLAWAY

Abstract. Document drafting is a central judicial problem-solving activity. Development of auto-

1 An illocutionary operator is a speech act such as informing, requesting, warning, or promising.

Section 2 describes a representative class of judicial documents - appellate

2. Judicial Document Drafting

Colorado Court of Appeals Order

Figure I. A typical show-cause order.

One approach to providing automated assistance in the drafting of show-cause

3. Document Drafting Using Discourse Structures

No.I7CAOSI4 "fr. O. No. ISCV269

Figure 2. The illocutionary and rhetorical structure of Canada

Cokndo CGIIn CIt AA-II Order

It f'Idw ~ rhI: ttriI il not. ftna.I)JdpmeD:

pronooncirW it to do in Older IO~J ddalnine the

RuII(AuthonI:y) . . orthe pM- irnoalved in the JIKlCOOdirw."

Figure 3. The illocutionary and rhetorical structure of Kirkpatrick

In summary, the discourse structure of a document consists of illocutionary

3.2. USING DOCUMENT GRAMMARS FOR DRAFTING

The illocutionary and rhetorical operators necessary to construct the discourse

• Analyzing a representative set of documents to determine their illocutionary

Kirkpatrick's illocutionary structure differs from that of Canada in several

EabIiIlfl(Pleqt-for-4i._O No. 9SCA0431 Tc. Ct. No. 91CVOSI

1b: HcdIert W. SmydIa. pro ...... _

ODderCUallmely-NOA-RMiDaae' Copiaco: c:o...lofRecurd

Figure 4. The illocutionary and rhetorical structure of Smythe

This example illustrates informally how a document grammar representing the

4. An Implemented Document Planner

4.1. FOUNDATIONS OF DISCOURSE GENERATION

The discourse-based approach to document planning builds on a strong founda-

4.2. REPRESENTING DOCUMENT PLANNING KNOWLEDGE

Schema-based approaches and their derivatives, such as explanation design

In addition to the illocutionary and rhetorical operators, the document grammar

specifies how achieving the illocutionary goals of establishing jurisdictional de-

4.3. DOCUMENT CREATION

(1) The Establishment Module

Figure 5. Operator modules of the DOCU-PLANNER's document grammar

(pattern (untimely-defect subject-defect finality-defect))

(pattern (appeal-filed filing-date end-sent)

Figure 6. Unification formalism of illocutionary and rhetorical operators

Case - - -___r.f Document Planner t-------....,

Figure 8. Representation of case facts of Smythe vs. Smythe

Figure 9. The illocutionary structure constructed for Smythe

is accomplished by attempting to apply an alternative illocutionary operator with

6. Discussion and Future Work

In this paper we have presented a model of the illocutionary and rhetorical

Establish(Preqs-For-Dismissal) ==> Establish(Jurisdictional-Defect)

Establish(JurisdictionaI-Defect) ==> Establish(Appeals)

Establish(Appeals) ==> One-Of [ Find(No-Main-Appeal),

Establish(Subject-Defect) ==> IF <subject-matter>